Mahi Luthra, Jiayi Shen, Maxime Poli, Angelo Ortiz Tandazo, Yosuke Higuchi, Youssef Benchekroun, Martin Gleize, Charles-Éric Saint-James, Dongyan Lin, Phillip Rust, Angel Villar-Corrales, Surya, Vanessa Stark, Rashel Moritz, Juan Pino, Yann LeCun, Emmanuel Dupoux. Proceedings of the 64th Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers). 2026.

A number of different architectures and loss functions have been applied to the problem of self-supervised learning (SSL), with the goal of developing embeddings that provide the best possible pre-training for as-yet-unknown, lightly supervised downstream tasks. One of these SSL criteria is to maximize the entropy of a set of embeddings in some compact space. But the goal of maximizing the embedding entropy often depends—whether explicitly or implicitly—upon high dimensional entropy estimates, which typically perform poorly in more than a few dimensions. In this paper, we motivate an effective entropy maximization criterion (E2MC), defined in terms of easy-to-estimate, low-dimensional constraints. We demonstrate that using it to continue training an already-trained SSL model for only a handful of epochs leads to a consistent and, in some cases, significant improvement in downstream performance. We perform careful ablation studies to show that the improved performance is due to the proposed add-on criterion. We also show that continued pre-training with alternative criteria does not lead to notable improvements, and in some cases, even degrades performance.

Normalization layers are ubiquitous in modern neural networks and have long been considered essential. This work demonstrates that Transformers without normalization can achieve the same or better performance using a remarkably simple technique. We introduce Dynamic Tanh (DyT), an element-wise operation DyT(x) = tanh(αx), as a dropin replacement for normalization layers in Transformers. DyT is inspired by the observation that layer normalization in Transformers often produces tanh-like, S-shaped input-output mappings. By incorporating DyT, Transformers without normalization can match or exceed the performance of their normalized counterparts, mostly without hyperparameter tuning. We validate the effectiveness of Transformers with DyT across diverse settings, from recognition to generation, supervised to self-supervised learning, and computer vision to language models. These findings challenge the conventional understanding that normalization layers are indispensable in modern neural networks, and offer new insights into their role in deep networks.

Navigation is a fundamental skill of agents with visual-motor capabilities. We introduce a Navigation World Model (NWM), a controllable video generation model that predicts future visual observations based on past observations and navigation actions. To capture complex environment dynamics, NWM employs a Conditional Diffusion Transformer (CDiT), trained on a diverse collection of egocentric videos of both human and robotic agents, and scaled up to 1 billion parameters. In familiar environments, NWM can plan navigation trajectories by simulating them and evaluating whether they achieve the desired goal. Unlike supervised navigation policies with fixed behavior, NWM can dynamically incorporate constraints during planning. Experiments demonstrate its effectiveness in planning trajectories from scratch or by ranking trajectories sampled from an external policy. Furthermore, NWM leverages its learned visual priors to imagine trajectories in unfamiliar environments from a single input image, making it a flexible and powerful tool for next-generation navigation systems1.

Vision-Based pose estimation of articulated robots with unknown joint angles has applications in collaborative robotics and human-robot interaction tasks. Current frameworks use neural network encoders to extract image features and downstream layers to predict joint angles and robot pose. While images of robots inherently contain rich information about the robot’s physical structures, existing methods often fail to leverage it fully; therefore, limiting performance under occlusions and truncations. To address this, we introduce RoboPEPP, a method that fuses information about the robot’s physical model into the encoder using a masking-based self-supervised embedding-predictive architecture. Specifically, we mask the robot’s joints and pre-train an encoder-predictor model to infer the joints’ embeddings from surrounding unmasked regions, enhancing the encoder’s understanding of the robot’s physical model. The pre-trained encoder-predictor pair, along with joint angle and keypoint prediction networks, is then fine-tuned for pose and joint angle estimation. Random masking of input during fine-tuning and keypoint filtering during evaluation further improves robustness. Our method, evaluated on several datasets, achieves the best results in robot pose and joint angle estimation while being the least sensitive to occlusions and requiring the lowest execution time. The code is available at https://github.com/raktimgg/RoboPEPP.

Accurate uncertainty estimation is crucial for deploying neural networks in risk-sensitive applications such as medical diagnosis. Monte Carlo Dropout is a widely used technique for approximating predictive uncertainty by performing stochastic forward passes with dropout during inference. However, using static dropout rates across all layers and inputs can lead to suboptimal uncertainty estimates, as it fails to adapt to the varying characteristics of individual inputs and network layers. Existing approaches optimize dropout rates during training using labeled data, resulting in fixed inference-time parameters that cannot adjust to new data distributions, compromising uncertainty estimates in Monte Carlo simulations.In this paper, we propose Rate-In, an algorithm that dynamically adjusts dropout rates during inference by quantifying the information loss induced by dropout in each layer’s feature maps. By treating dropout as controlled noise injection and leveraging information-theoretic principles, Rate-In adapts dropout rates per layer and per input instance without requiring ground truth labels. By quantifying the functional information loss in feature maps, we adaptively tune dropout rates to maintain perceptual quality across diverse medical imaging tasks and architectural configurations. Our extensive empirical study on synthetic data and real-world medical imaging tasks demonstrates that Rate-In improves calibration and sharpens uncertainty estimates compared to fixed or heuristic dropout rates without compromising predictive performance. Rate-In offers a practical, unsupervised, inference-time approach to optimizing dropout for more reliable predictive uncertainty estimation in critical applications.

Transformer encoders are critical for a wide range of Natural Language Processing (NLP) tasks, yet their compute-efficiency remains poorly understood.We present the first comprehensive empirical investigation of compute-optimal pretraining for encoder transformers using the Masked Language Modeling (MLM) objective.Across hundreds of carefully controlled runs we vary model size, data size, batch size, learning rate, and masking ratio, with increasing compute budget.The compute-optimal data-to-model ratio of Transformer encoder models is 10 to 100 times larger than the ratio of auto-regressive models.Using these recipes, we train OptiBERT, a family of compute-optimal BERT-style models that matches or surpasses leading baselines-including ModernBERT and NeoBERT-on GLUE and MTEB while training with dramatically less FLOPS.

Visual Self-Supervised Learning (SSL) currently underperforms Contrastive Language-Image Pretraining (CLIP) in multimodal settings such as Visual Question Answering (VQA). This multimodal gap is often attributed to the semantics introduced by language supervision, even though visual SSL and CLIP models are often trained on different data. In this work, we ask the question: "Do visual self-supervised approaches lag behind CLIP due to the lack of language supervision, or differences in the training data?" We study this question by training both visual SSL and CLIP models on the same MetaCLIP data, and leveraging VQA as a diverse testbed for vision encoders. In this controlled setup, visual SSL models scale better than CLIP models in terms of data and model capacity, and visual SSL performance does not saturate even after scaling up to 7B parameters. Consequently, we observe visual SSL methods achieve CLIP-level performance on a wide range of VQA and classic vision benchmarks. These findings demonstrate that pure visual SSL can match language-supervised visual pretraining at scale, opening new opportunities for vision-centric representation learning.

In this work, we propose Visual-Predictive Instruction Tuning (VPiT) - a simple and effective extension to visual instruction tuning that enables a pretrained LLM to quickly morph into an unified autoregressive model capable of generating both text and visual tokens. VPiT teaches an LLM to predict discrete text tokens and continuous visual tokens from any input sequence of image and text data curated in an instruction-following format. Our empirical investigation reveals several intriguing properties of VPiT: (1) visual generation ability emerges as a natural byproduct of improved visual understanding, and can be unlocked efficiently with a small amount of generation data; (2) while we find understanding and generation to be mutually beneficial, understanding data contributes to both capabilities more effectively than generation data. Building upon these findings, we train our MetaMorph model and achieve competitive performance on both visual understanding and generation. In visual generation, MetaMorph can leverage the world knowledge and reasoning abilities gained from LLM pretraining, and overcome common failure modes exhibited by other generation models. Our results suggest that LLMs may have strong "prior" vision capabilities that can be efficiently adapted to both visual understanding and generation with a relatively simple instruction tuning process.

Whole-body control for humanoids is challenging due to the high-dimensional nature of the problem, coupled with the inherent instability of a bipedal morphology. Learning from visual observations further exacerbates this difficulty. In this work, we explore highly data-driven approaches to visual whole-body humanoid control based on reinforcement learning, without any simplifying assumptions, reward design, or skill primitives. Specifically, we propose a hierarchical world model in which a high-level agent generates commands based on visual observations for a low-level agent to execute, both of which are trained with rewards. Our approach produces highly performant control policies in 8 tasks with a simulated 56-DoF humanoid, while synthesizing motions that are broadly preferred by humans. Code and videos: https://www.nicklashansen.com/rlpuppeteer

Decoder-only Transformers often struggle with complex reasoning tasks, particularly arithmetic reasoning requiring multiple sequential operations. In this work, we identify representation collapse in the model’s intermediate layers as a key factor limiting their reasoning capabilities. To address this, we propose Sequential Variance-Covariance Regularization (Seq-VCR), which enhances the entropy of intermediate representations and prevents collapse. Combined with dummy pause tokens as substitutes for chain-of-thought (CoT) tokens, our method significantly improves performance in arithmetic reasoning problems. In the challenging 5 × 5 integer multiplication task, our approach achieves 99.5% exact match accuracy, outperforming models of the same size (which yield 0% accuracy) and GPT-4 with five-shot CoT prompting (44%). We also demonstrate superior results on arithmetic expression and longest increasing subsequence (LIS) datasets. Our findings highlight the importance of preventing intermediate layer representation collapse to enhance the reasoning capabilities of Transformers and show that Seq-VCR offers an effective solution without requiring explicit CoT supervision.

Self-supervised learning has driven significant progress in learning from single-subject, _iconic_ images. However, there are still unanswered questions about the use of minimally-curated, naturalistic video data, which contain _dense_ scenes with many independent objects, imbalanced class distributions, and varying object sizes. In this paper, we propose PooDLe, a self-supervised learning method that combines an invariance-based objective on pooled representations with a dense SSL objective that enforces equivariance to optical flow warping. Our results show that a unified objective applied at multiple feature scales is essential for learning effective image representations from naturalistic videos. We validate our method with experiments on the BDD100K driving video dataset and the Walking Tours first-person video dataset, demonstrating its ability to capture spatial understanding from a dense objective and semantic understanding via a pooled representation objective.

Test set contamination, wherein test data from a benchmark ends up in a newer model's training set, is a well-documented obstacle for fair LLM evaluation and can quickly render benchmarks obsolete. To mitigate this, many recent benchmarks crowdsource new prompts and evaluations from human or LLM judges; however, these can introduce significant biases, and break down when scoring hard questions. In this work, we introduce a new benchmark for LLMs designed to be resistant to both test set contamination and the pitfalls of LLM judging and human crowdsourcing. We release LiveBench, the first benchmark that (1) contains frequently-updated questions from recent information sources, (2) scores answers automatically according to objective ground-truth values, and (3) contains a wide variety of challenging tasks, spanning math, coding, reasoning, language, instruction following, and data analysis. To achieve this, LiveBench contains questions that are based on recently-released math competitions, arXiv papers, news articles, and datasets, and it contains harder, contamination-limited versions of tasks from previous benchmarks such as Big-Bench Hard, AMPS, and IFEval. We evaluate many prominent closed-source models, as well as dozens of open-source models ranging from 0.5B to 405B in size. LiveBench is difficult, with top models achieving below 70% accuracy. We release all questions, code, and model answers. Questions are added and updated on a monthly basis, and we release new tasks and harder versions of tasks over time so that LiveBench can distinguish between the capabilities of LLMs as they improve in the future. We welcome community engagement and collaboration for expanding the benchmark tasks and models.

Unsupervised representation learning has seen tremendous progress. However, it is constrained by its reliance on domain specific stationarity and topology, a limitation not found in biological intelligence systems. For instance, unlike computer vision, human vision can process visual signals sampled from highly irregular and non-stationary sensors. We introduce a novel framework that learns from high-dimensional data without prior knowledge of stationarity and topology. Our model, abbreviated as URLOST, combines a learnable self-organizing layer, spectral clustering, and a masked autoencoder (MAE). We evaluate its effectiveness on three diverse data modalities including simulated biological vision data, neural recordings from the primary visual cortex, and gene expressions. Compared to state-of-the-art unsupervised learning methods like SimCLR and MAE, our model excels at learning meaningful representations across diverse modalities without knowing their stationarity or topology. It also outperforms other methods that are not dependent on these factors, setting a new benchmark in the field. We position this work as a step toward unsupervised learning methods capable of generalizing across diverse high-dimensional data modalities.

From extracting features to generating text, the outputs of large language models (LLMs) typically rely on their final layers, following the conventional wisdom that earlier layers capture only low-level cues. However, our analysis shows that intermediate layers can encode even richer representations, often improving performance on a wide range of downstream tasks. To explain and quantify these hidden-layer properties, we propose a unified framework of representation quality metrics based on information theory, geometry, and invariance to input perturbations. Our framework highlights how each model layer balances information compression and signal preservation, revealing why mid-depth embeddings can exceed the last layer’s performance. Through extensive experiments on 32 text-embedding tasks across various architectures (transformers, state-space models) and domains (language, vision), we demonstrate that intermediate layers consistently provide stronger features, challenging the standard view on final-layer embeddings and opening new directions on using mid-layer representations for more robust and accurate representations.

The ability to predict future outcomes given control actions is fundamental for physical reasoning. However, such predictive models, often called world models, remain challenging to learn and are typically developed for task-specific solutions with online policy learning. To unlock world models' true potential, we argue that they should 1) be trainable on offline, pre-collected trajectories, 2) support test-time behavior optimization, and 3) facilitate task-agnostic reasoning. To this end, we present DINO World Model (DINO-WM), a new method to model visual dynamics without reconstructing the visual world. DINO-WM leverages spatial patch features pre-trained with DINOv2, enabling it to learn from offline behavioral trajectories by predicting future patch features. This allows DINO-WM to achieve observational goals through action sequence optimization, facilitating task-agnostic planning by treating goal features as prediction targets. We demonstrate that DINO-WM achieves zero-shot behavioral solutions at test time on six environments without expert demonstrations, reward modeling, or pre-learned inverse models, outperforming prior state-of-the-art work across diverse task families such as arbitrarily configured mazes, push manipulation with varied object shapes, and multi-particle scenarios.

We train models to predict ego-centric video from human actions (PEVA), given the past video and an action represented by the relative 3D body pose. By conditioning on kinematic pose trajectories, structured by the joint hierarchy of the body, our model learns to simulate how physical human actions shape the environment from a first-person point of view. We train an auto-regressive conditional diffusion transformer on Nymeria, a large-scale dataset of real-world egocentric video and body pose capture. We further design a hierarchical evaluation protocol with increasingly challenging tasks, enabling a comprehensive analysis of the model’s embodied prediction and control abilities. Our work represents an initial attempt to tackle the challenges of modeling complex real-world environments and embodied agent behaviors with video prediction from the perspective of a human.

Visual imitation learning enables robotic agents to acquire skills by observing expert demonstration videos. In the one-shot setting, the agent generates a policy after observing a single expert demonstration without additional fine-tuning. Existing approaches typically train and evaluate on the same set of tasks, varying only object configurations, and struggle to generalize to unseen tasks with different semantic or structural requirements. While some recent methods attempt to address this, they exhibit low success rates on hard test tasks that, despite being visually similar to some training tasks, differ in context and require distinct responses. Additionally, most existing methods lack an explicit model of environment dynamics, limiting their ability to reason about future states. To address these limitations, we propose a novel framework for one-shot visual imitation learning via world-model-guided trajectory generation. Given an expert demonstration video and the agent’s initial observation, our method leverages a learned world model to predict a sequence of latent states and actions. This latent trajectory is then decoded into physical waypoints that guide the agent’s execution. Our method is evaluated on two simulated benchmarks and three real-world robotic platforms, where it consistently outperforms prior approaches, with over 30% improvement in some cases.

A long-standing goal in AI is to develop agents capable of solving diverse tasks across a range of environments, including those never seen during training. Two dominant paradigms address this challenge: (i) reinforcement learning (RL), which learns policies via trial and error, and (ii) optimal control, which plans actions using a known or learned dynamics model. However, their comparative strengths in the offline setting—where agents must learn from reward-free trajectories—remain underexplored. In this work, we systematically evaluate RL and control-based methods on a suite of navigation tasks, using offline datasets of varying quality. On the RL side, we consider goal-conditioned and zero-shot methods. On the control side, we train a latent dynamics model using the Joint Embedding Predictive Architecture (JEPA) and employ it for planning. We investigate how factors such as data diversity, trajectory quality, and environment variability influence the performance of these approaches. Our results show that model-free RL benefits most from large amounts of high-quality data, whereas model-based planning generalizes better to unseen layouts and is more data-efficient, while achieving trajectory stitching performance comparable to leading model-free methods. Notably, planning with a latent dynamics model proves to be a strong approach for handling suboptimal offline data and adapting to diverse environments.

Is vision good enough for language? Recent advancements in multimodal models primarily stem from the powerful reasoning abilities of large language models (LLMs). However, the visual component typically depends only on the instance-level contrastive language-image pre-training (CLIP). Our research reveals that the visual capabilities in recent MultiModal LLMs (MLLMs) still exhibit systematic shortcomings. To understand the roots of these errors, we explore the gap between the visual embedding space of CLIP and vision-only self-supervised learning. We identify “CLIP-blind pairs”- images that CLIP perceives as similar despite their clear visual differences. With these pairs, we construct the Multimodal Visual Patterns (MMVP) benchmark. MMVP exposes areas where state-of-the-art systems, including GPT-4V, struggle with straightforward questions across nine basic visual patterns, often providing incorrect answers and hallucinated explanations. We further evaluate various CLIP-based vision-and-language models and found a notable correlation between visual patterns that challenge CLIP models and those problematic for multimodal LLMs. As an initial effort to address these issues, we propose a Mixture of Features (MoF) approach, demonstrating that integrating vision self-supervised learning features with MLLMs can significantly enhance their visual grounding capabilities. Together, our research suggests visual representation learning remains an open challenge, and accurate visual grounding is crucial for future successful multimodal systems.

Representation learning on text-attributed graphs (TAGs) has become a critical research problem in recent years. A typical example of a TAG is a paper citation graph, where the text of each paper serves as node attributes. Initial graph neural network (GNN) pipelines handled these text attributes by transforming them into shallow or hand-crafted features, such as skip-gram or bag-of-words features. Recent efforts have focused on enhancing these pipelines with language models (LMs), which typically demand intricate designs and substantial computational resources. With the advent of powerful large language models (LLMs) such as GPT or Llama2, which demonstrate an ability to reason and to utilize general knowledge, there is a growing need for techniques which combine the textual modelling abilities of LLMs with the structural learning capabilities of GNNs. Hence, in this work, we focus on leveraging LLMs to capture textual information as features, which can be used to boost GNN performance on downstream tasks. A key innovation is our use of \emph{explanations as features}: we prompt an LLM to perform zero-shot classification, request textual explanations for its decision-making process, and design an \emph{LLM-to-LM interpreter} to translate these explanations into informative features for downstream GNNs. Our experiments demonstrate that our method achieves state-of-the-art results on well-established TAG datasets, including \texttt{Cora}, \texttt{PubMed}, \texttt{ogbn-arxiv}, as well as our newly introduced dataset, \texttt{tape-arxiv23}. Furthermore, our method significantly speeds up training, achieving a 2.88 times improvement over the closest baseline on \texttt{ogbn-arxiv}. Lastly, we believe the versatility of the proposed method extends beyond TAGs and holds the potential to enhance other tasks involving graph-text data~\footnote{Our codes and datasets are available at: \url{https://github.com/XiaoxinHe/TAPE}}.

We introduce GAIA, a benchmark for General AI Assistants that, if solved, would represent a milestone in AI research. GAIA proposes real-world questions that require a set of fundamental abilities such as reasoning, multi-modality handling, web browsing, and generally tool-use proficiency. GAIA questions are conceptually simple for humans yet challenging for most advanced AIs: we show that human respondents obtain 92% vs. 15% for GPT-4 equipped with plugins. This notable performance disparity contrasts with the recent trend of LLMs outperforming humans on tasks requiring professional skills in e.g. law or chemistry. GAIA’s philosophy departs from the current trend in AI benchmarks suggesting to target tasks that are ever more difficult for humans. We posit that the advent of Artificial General Intelligence (AGI) hinges on a system’s capability to exhibit similar robustness as the average human does on such questions. Using GAIA’s methodology, we devise 466 questions and their answer. We release our questions while retaining answers to 300 of them to power a leader-board accessible at https://huggingface.co/gaia-benchmark.

Input space reconstruction is an attractive representation learning paradigm. Despite interpretability benefit of reconstruction and generation, we identify a misalignment between learning to reconstruct, and learning for perception. We show that the former allocates a model's capacity towards a subspace of the data explaining the observed variance--a subspace with uninformative features for the latter. For example, the supervised TinyImagenet task with images projected onto the top subspace explaining 90% of the pixel variance can be solved with 45% test accuracy. Using the bottom subspace instead, accounting for only 20% of the pixel variance, reaches 55% test accuracy. Learning by reconstruction is also wasteful as the features for perception are learned last, pushing the need for long training schedules. We finally prove that learning by denoising can alleviate that misalignment for some noise strategies, e.g., masking. While tuning the noise strategy without knowledge of the perception task seems challenging, we provide a solution to detect if a noise strategy is never beneficial regardless of the perception task, e.g., additive Gaussian noise.

Masked Image Modeling (MIM) is a promising self-supervised learning approach that enables learning from unlabeled images. Despite its recent success, learning good representations through MIM remains challenging because it requires predicting the right semantic content in accurate locations. For example, given an incomplete picture of a dog, we can guess that there is a tail, but we cannot determine its exact location. In this work, we propose to incorporate location uncertainty to MIM by using stochastic positional embeddings (StoP). Specifically, we condition the model on stochastic masked token positions drawn from a gaussian distribution. We show that using StoP reduces overfitting to location features and guides the model toward learning features that are more robust to location uncertainties. Quantitatively, using StoP improves downstream MIM performance on a variety of downstream tasks. For example, linear probing on ImageNet using ViT-B is improved by $+1.7\%$, and by $2.5\%$ for ViT-H using 1% of the data.

Entropy minimization (EM) is frequently used to increase the accuracy of classification models when they're faced with new data at test time. EM is a self-supervised learning method that optimizes classifiers to assign even higher probabilities to their top predicted classes. In this paper, we analyze why EM works when adapting a model for a few steps and why it eventually fails after adapting for many steps. We show that, at first, EM causes the model to embed test images close to training images, thereby increasing model accuracy. After many steps of optimization, EM makes the model embed test images far away from the embeddings of training images, which results in a degradation of accuracy. Building upon our insights, we present a method for solving a practical problem: estimating a model's accuracy on a given arbitrary dataset without having access to its labels. Our method estimates accuracy by looking at how the embeddings of input images change as the model is optimized to minimize entropy. Experiments on 23 challenging datasets show that our method sets the SoTA with a mean absolute error of 5.75%, an improvement of 29.62% over the previous SoTA on this task. Our code is available at: https://github.com/oripress/EntropyEnigma

Given a graph with textual attributes, we enable users to `chat with their graph': that is, to ask questions about the graph using a conversational interface. In response to a user's questions, our method provides textual replies and highlights the relevant parts of the graph. While existing works integrate large language models (LLMs) and graph neural networks (GNNs) in various ways, they mostly focus on either conventional graph tasks (such as node, edge, and graph classification), or on answering simple graph queries on small or synthetic graphs. In contrast, we develop a flexible question-answering framework targeting real-world textual graphs, applicable to multiple applications including scene graph understanding, common sense reasoning, and knowledge graph reasoning. Toward this goal, we first develop a Graph Question Answering (GraphQA) benchmark with data collected from different tasks. Then, we propose our \textit{G-Retriever} method, introducing the first retrieval-augmented generation (RAG) approach for general textual graphs, which can be fine-tuned to enhance graph understanding via soft prompting. To resist hallucination and to allow for textual graphs that greatly exceed the LLM's context window size, \textit{G-Retriever} performs RAG over a graph by formulating this task as a Prize-Collecting Steiner Tree optimization problem. Empirical evaluations show that our method outperforms baselines on textual graph tasks from multiple domains, scales well with larger graph sizes, and mitigates hallucination.~\footnote{Our codes and datasets are available at: \url{https://github.com/XiaoxinHe/G-Retriever}}

We introduce Cambrian-1, a family of multimodal LLMs (MLLMs) designed with a vision-centric approach. While stronger language models can enhance multimodal capabilities, the design choices for vision components are often insufficiently explored and disconnected from visual representation learning research. This gap hinders accurate sensory grounding in real-world scenarios. Our study uses LLMs and visual instruction tuning as an interface to evaluate various visual representations, offering new insights into different models and architectures—self-supervised, strongly supervised, or combinations thereof—based on experiments with over 15 vision models. We critically examine existing MLLM benchmarks, addressing the difficulties involved in consolidating and interpreting results from various tasks. To further improve visual grounding, we propose spatial vision aggregator (SVA), a dynamic and spatially-aware connector that integrates vision features with LLMs while reducing the number of tokens. Additionally, we discuss the curation of high-quality visual instruction-tuning data from publicly available sources, emphasizing the importance of distribution balancing. Collectively, Cambrian-1 not only achieves state-of-the-art performances but also serves as a comprehensive, open cookbook for instruction-tuned MLLMs. We provide model weights, code, supporting tools, datasets, and detailed instruction-tuning and evaluation recipes. We hope our release will inspire and accelerate advancements in multimodal systems and visual representation learning.

Large vision-language models (VLMs) fine-tuned on specialized visual instruction-following data have exhibited impressive language reasoning capabilities across various scenarios. However, this fine-tuning paradigm may not be able to efficiently learn optimal decision-making agents in multi-step goal-directed tasks from interactive environments. To address this challenge, we propose an algorithmic framework that fine-tunes VLMs with reinforcement learning (RL). Specifically, our framework provides a task description and then prompts the VLM to generate chain-of-thought (CoT) reasoning, enabling the VLM to efficiently explore intermediate reasoning steps that lead to the final text-based action. Next, the open-ended text output is parsed into an executable action to interact with the environment to obtain goal-directed task rewards. Finally, our framework uses these task rewards to fine-tune the entire VLM with RL. Empirically, we demonstrate that our proposed framework enhances the decision-making capabilities of VLM agents across various tasks, enabling 7b models to outperform commercial models such as GPT4-V or Gemini. Furthermore, we find that CoT reasoning is a crucial component for performance improvement, as removing the CoT reasoning results in a significant decrease in the overall performance of our method.

Undoubtedly, Deep Neural Networks (DNNs), from AlexNet to ResNet to Transformer, have sparked revolutionary advancements in diverse computer vision tasks. The scale of DNNs has grown exponentially due to the rapid development of computational resources. Despite the tremendous success, DNNs typically depend on massive amounts of training data (especially the recent various foundation models) to achieve high performance and are brittle in that their performance can degrade severely with small changes in their operating environment. Generally, collecting massive-scale training datasets is costly or even infeasible, as for certain fields, only very limited or no examples at all can be gathered. Nevertheless, collecting, labeling, and vetting massive amounts of practical training data is certainly difficult and expensive, as it requires the painstaking efforts of experienced human annotators or experts, and in many cases, prohibitively costly or impossible due to some reason, such as privacy, safety or ethic issues.

This paper demonstrates an approach for learning highly semantic image representations without relying on hand-crafted data-augmentations. We introduce the Image-based Joint-Embedding Predictive Architecture (I-JEPA), a non-generative approach for self-supervised learning from images. The idea behind I-JEPA is simple: from a single context block, predict the representations of various target blocks in the same image. A core design choice to guide I-JEPA towards producing semantic representations is the masking strategy; specifically, it is crucial to (a) sample target blocks with sufficiently large scale (semantic), and to (b) use a sufficiently informative (spatially distributed) context block. Empirically, when combined with Vision Transformers, we find I-JEPA to be highly scalable. For instance, we train a ViT-Huge/14 on ImageNet using 16 A100 GPUs in under 72 hours to achieve strong downstream performance across a wide range of tasks, from linear classification to object counting and depth prediction.

One fruitful formulation of Deep Networks (DNs) enabling their theoretical study and providing practical guidelines to practitioners relies on Piecewise Affine Splines. In that realm, a DN’s input-mapping is expressed as per-region affine map-ping where those regions are implicitly determined by the model’s architecture and form a partition of their input space. That partition –which is involved in all the results spanned from this line of research– has so far only been computed on 2/3-dimensional slices of the DN’s input space or estimated by random sampling. In this paper, we provide the first parallel algorithm that does exact enumeration of the DN’s partition regions. The proposed algorithm enables one to finally assess the closeness of the commonly employed approximations methods, e.g. based on random sampling of the DN input space. One of our key finding is that if one is only interested in regions with "large" volume, then uniform sampling of the space is highly efficient, but that if one is also interested in discovering the "small" regions of the partition, then uniform sampling is exponentially costly with the DN’s input space dimension. On the other hand, our proposed method has complexity scaling linearly with input dimension and the number of regions.

Deep Neural Networks (DNNs) outshine alternative function approximators in many settings thanks to their modularity in composing any desired differentiable operator. The formed parametrized functional is then tuned to solve a task at hand from simple gradient descent. This modularity comes at the cost of making strict enforcement of constraints on DNNs, e.g. from a priori knowledge of the task, or from desired physical properties, an open challenge. In this paper we propose the first provable affine constraint enforcement method for DNNs that only requires minimal changes into a given DNN’s forward-pass, that is computationally friendly, and that leaves the optimization of the DNN’s parameter to be unconstrained, i.e. standard gradient-based method can be employed. Our method does not require any sampling and provably ensures that the DNN fulfills the affine constraint on a given input space’s region at any point during training, and testing. We coin this method POLICE, standing for Provably Optimal LInear Constraint Enforcement. Github: https://github.com/RandallBalestriero/POLICE

Self-Supervised Learning (SSL) has emerged as the solution of choice to learn transferable representations from unlabeled data. However, SSL requires to build samples that are known to be semantically akin, i.e. positive views. Requiring such knowledge is the main limitation of SSL and is often tackled by ad-hoc strategies e.g. applying known data-augmentations to the same input. In this work, we formalize and generalize this principle through Positive Active Learning (PAL) where an oracle queries semantic relationships between samples. PAL achieves three main objectives. First, it unveils a theoretically grounded learning framework beyond SSL, based on similarity graphs, that can be extended to tackle supervised and semi-supervised learning depending on the employed oracle. Second, it provides a consistent algorithm to embed a priori knowledge, e.g. some observed labels, into any SSL losses without any change in the training pipeline. Third, it provides a proper active learning framework yielding low-cost solutions to annotate datasets, arguably bringing the gap between theory and practice of active learning that is based on simple-to-answer-by-non-experts queries of semantic relationships between inputs.

Self-supervised learning (SSL) has emerged as a powerful framework to learn representations from raw data without supervision. Yet in practice, engineers face issues such as instability in tuning optimizers and collapse of representations during training. Such challenges motivate the need for a theory to shed light on the complex interplay between the choice of data augmentation, network architecture, and training algorithm. % on the resulting performance in downstream tasks. We study such an interplay with a precise analysis of generalization performance on both pretraining and downstream tasks in kernel regimes, and highlight several insights for SSL practitioners that arise from our theory.

Joint-Embedding Self Supervised Learning (JE-SSL) has seen a rapid development, with the emergence of many method variations but only few principled guidelines that would help practitioners to successfully deploy them. The main reason for that pitfall comes from JE-SSL’s core principle of not employing any input reconstruction therefore lacking visual cues of unsuccessful training. Adding non informative loss values to that, it becomes difficult to deploy SSL on a new dataset for which no labels can help to judge the quality of the learned representation. In this study, we develop a simple unsupervised criterion that is indicative of the quality of the learned JE-SSL representations: their effective rank. Albeit simple and computationally friendly, this method —coined RankMe— allows one to assess the performance of JE-SSL representations, even on different downstream datasets, without requiring any labels. A further benefit of RankMe is that it does not have any training or hyper-parameters to tune. Through thorough empirical experiments involving hundreds of training episodes, we demonstrate how RankMe can be used for hyperparameter selection with nearly no reduction in final performance compared to the current selection method that involve a dataset’s labels. We hope that RankMe will facilitate the deployment of JE-SSL towards domains that do not have the opportunity to rely on labels for representations’ quality assessment.

Recent progress has been made towards learning invariant or equivariant representations with self-supervised learning. While invariant methods are evaluated on large scale datasets, equivariant ones are evaluated in smaller, more controlled, settings. We aim at bridging the gap between the two in order to learn more diverse representations that are suitable for a wide range of tasks. We start by introducing a dataset called 3DIEBench, consisting of renderings from 3D models over 55 classes and more than 2.5 million images where we have full control on the transformations applied to the objects. We further introduce a predictor architecture based on hypernetworks to learn equivariant representations with no possible collapse to invariance. We introduce SIE (Split Invariant-Equivariant) which combines the hypernetwork-based predictor with representations split in two parts, one invariant, the other equivariant, to learn richer representations. We demonstrate significant performance gains over existing methods on equivariance related tasks from both a qualitative and quantitative point of view. We further analyze our introduced predictor and show how it steers the learned latent space. We hope that both our introduced dataset and approach will enable learning richer representations without supervision in more complex scenarios. Code and data are available at https://github.com/garridoq/SIE.

Graph Neural Networks (GNNs) have shown great potential in the field of graph representation learning. Standard GNNs define a local message-passing mechanism which propagates information over the whole graph domain by stacking multiple layers. This paradigm suffers from two major limitations, over-squashing and poor long-range dependencies, that can be solved using global attention but significantly increases the computational cost to quadratic complexity. In this work, we propose an alternative approach to overcome these structural limitations by leveraging the ViT/MLP-Mixer architectures introduced in computer vision. We introduce a new class of GNNs, called Graph ViT/MLP-Mixer, that holds three key properties. First, they capture long-range dependency and mitigate the issue of over-squashing as demonstrated on Long Range Graph Benchmark and TreeNeighbourMatch datasets. Second, they offer better speed and memory efficiency with a complexity linear to the number of nodes and edges, surpassing the related Graph Transformer and expressive GNN models. Third, they show high expressivity in terms of graph isomorphism as they can distinguish at least 3-WL non-isomorphic graphs. We test our architecture on 4 simulated datasets and 7 real-world benchmarks, and show highly competitive results on all of them. The source code is available for reproducibility at: https://github.com/XiaoxinHe/Graph-ViT-MLPMixer.

Understanding the learned representation and underlying mechanisms of Self-Supervised Learning (SSL) often poses a challenge. In this paper, we ‘reverse engineer’ SSL, conducting an in-depth empirical analysis of its learned internal representations, encompassing diverse models, architectures, and hyperparameters. Our study reveals an intriguing process within the SSL training: an inherent facilitation of semantic label-based clustering, which is surprisingly driven by the regularization component of the SSL objective. This clustering not only enhances downstream classification, but also compresses the information. We further illustrate that the alignment of the SSL-trained representation is more pronounced with semantic classes rather than random functions. Remarkably, the learned representations align with semantic classes across various hierarchical levels, with this alignment intensifying when going deeper into the network. This ‘reverse engineering’ approach provides valuable insights into the inner mechanism of SSL and their influences on the performance across different class sets.

Machine learning for differential equations paves the way for computationally efficient alternatives to numerical solvers, with potentially broad impacts in science and engineering. Though current algorithms typically require simulated training data tailored to a given setting, one may instead wish to learn useful information from heterogeneous sources, or from real dynamical systems observations that are messy or incomplete. In this work, we learn general-purpose representations of PDEs from heterogeneous data by implementing joint embedding methods for self-supervised learning (SSL), a framework for unsupervised representation learning that has had notable success in computer vision. Our representation outperforms baseline approaches to invariant tasks, such as regressing the coefficients of a PDE, while also improving the time-stepping performance of neural solvers. We hope that our proposed methodology will prove useful in the eventual development of general-purpose foundation models for PDEs.

Variance-Invariance-Covariance Regularization (VICReg) is a self-supervised learning (SSL) method that has shown promising results on a variety of tasks. However, the fundamental mechanisms underlying VICReg remain unexplored. In this paper, we present an information-theoretic perspective on the VICReg objective. We begin by deriving information-theoretic quantities for deterministic networks as an alternative to unrealistic stochastic network assumptions. We then relate the optimization of the VICReg objective to mutual information optimization, highlighting underlying assumptions and facilitating a constructive comparison with other SSL algorithms and derive a generalization bound for VICReg, revealing its inherent advantages for downstream tasks. Building on these results, we introduce a family of SSL methods derived from information-theoretic principles that outperform existing SSL techniques.

Deep learning has achieved tremendous success by training increasingly large models, which are then compressed for practical deployment. We propose a drastically different approach to compact and optimal deep learning: We decouple the Degrees of freedom (DoF) and the actual number of parameters of a model, optimize a small DoF with predefined random linear constraints for a large model of an arbitrary architecture, in one-stage end-to-end learning.Specifically, we create a recurrent parameter generator (RPG), which repeatedly fetches parameters from a ring and unpacks them onto a large model with random permutation and sign flipping to promote parameter decorrelation. We show that gradient descent can automatically find the best model under constraints with in fact faster convergence.Our extensive experimentation reveals a log-linear relationship between model DoF and accuracy. Our RPG demonstrates remarkable DoF reduction, and can be further pruned and quantized for additional run-time performance gain. For example, in terms of top-1 accuracy on ImageNet, RPG achieves 96% of ResNet18’s performance with only 18% DoF (the equivalent of one convolutional layer) and 52% of ResNet34’s performance with only 0.25% DoF! Our work shows significant potential of constrained neural opti-mization in compact and optimal deep learning.

Regularization is a fundamental technique to prevent over-fitting and to improve generalization performances by constraining a model's complexity. Current Deep Networks heavily rely on regularizers such as Data-Augmentation (DA) or weight-decay, and employ structural risk minimization, i.e. cross-validation, to select the optimal regularization hyper-parameters. In this study, we demonstrate that techniques such as DA or weight decay produce a model with a reduced complexity that is unfair across classes. The optimal amount of DA or weight decay found from cross-validation over all classes leads to disastrous model performances on some classes e.g. on Imagenet with a resnet50, the ``barn spider'' classification test accuracy falls from $68\%$ to $46\%$ only by introducing random crop DA during training. Even more surprising, such performance drop also appears when introducing uninformative regularization techniques such as weight decay. Those results demonstrate that our search for ever increasing generalization performance ---averaged over all classes and samples--- has left us with models and regularizers that silently sacrifice performances on some classes. This scenario can become dangerous when deploying a model on downstream tasks e.g. an Imagenet pre-trained resnet50 deployed on INaturalist sees its performances fall from $70\%$ to $30\%$ on class \#8889 when introducing random crop DA during the Imagenet pre-training phase. Those results demonstrate that finding a correct measure of a model's complexity without class-dependent preference remains an open research question.

Self-Supervised Learning (SSL) surmises that inputs and pairwise positive relationships are enough to learn meaningful representations. Although SSL has recently reached a milestone: outperforming supervised methods in many modalities\dots the theoretical foundations are limited, method-specific, and fail to provide principled design guidelines to practitioners. In this paper, we propose a unifying framework under the helm of spectral manifold learning. Through the course of this study, we will demonstrate that VICReg, SimCLR, BarlowTwins et al. correspond to eponymous spectral methods such as Laplacian Eigenmaps, ISOMAP et al.From this unified viewpoint, we obtain (i) the close-form optimal representation, (ii) the close-form optimal network parameters in the linear regime, (iii) the impact of the pairwise relations used during training on each of those quantities and on downstream task performances, and most importantly, (iv) the first theoretical bridge between contrastive and non-contrastive methods to global and local spectral methods respectively hinting at the benefits and limitations of each. For example, if the pairwise relation is aligned with the downstream task, all SSL methods produce optimal representations for that downstream task.

Data-Augmentation (DA) is known to improve performance across tasks and datasets. We propose a method to theoretically analyze the effect of DA and study questions such as: how many augmented samples are needed to correctly estimate the information encoded by that DA? How does the augmentation policy impact the final parameters of a model? We derive several quantities in close-form, such as the expectation and variance of an image, loss, and model's output under a given DA distribution. Up to our knowledge, we obtain the first explicit regularizer that corresponds to using DA during training for non-trivial transformations such as affine transformations, color jittering, or Gaussian blur. Those derivations open new avenues to quantify the benefits and limitations of DA. For example, given a loss at hand, we find that common DAs require tens of thousands of samples for the loss to be correctly estimated and for the model training to converge. We then show that for a training loss to have reduced variance under DA sampling, the model's saliency map (gradient of the loss with respect to the model's input) must align with the smallest eigenvector of the sample's covariance matrix under the considered DA augmentation; this is exactly the quantity estimated and regularized by TangentProp. Those findings also hint at a possible explanation on why models tend to shift their focus from edges to textures when specific DAs are employed.

Most recent self-supervised methods for learning image representations focus on either producing a global feature with invariance properties, or producing a set of local features. The former works best for classification tasks while the latter is best for detection and segmentation tasks. This paper explores the fundamental trade-off between learning local and global features. A new method called VICRegL is proposed that learns good global and local features simultaneously, yielding excellent performance on detection and segmentation tasks while maintaining good performance on classification tasks. Concretely, two identical branches of a standard convolutional net architecture are fed two differently distorted versions of the same image. The VICReg criterion is applied to pairs of global feature vectors. Simultaneously, the VICReg criterion is applied to pairs of local feature vectors occurring before the last pooling layer. Two local feature vectors are attracted to each other if their l2-distance is below a threshold or if their relative locations are consistent with a known geometric transformation between the two input images. We demonstrate strong performance on linear classification and segmentation transfer tasks. Code and pretrained models are publicly available at: https://github.com/facebookresearch/VICRegL

Vision-language (VL) pre-training has recently received considerable attention. However, most existing end-to-end pre-training approaches either only aim to tackle VL tasks such as image-text retrieval, visual question answering (VQA) and image captioning that test high-level understanding of images, or only target region-level understanding for tasks such as phrase grounding and object detection. We present FIBER (Fusion-In-the-Backbone-based transformER), a new VL model architecture that can seamlessly handle both these types of tasks. Instead of having dedicated transformer layers for fusion after the uni-modal backbones, FIBER pushes multimodal fusion deep into the model by inserting cross-attention into the image and text backbones to better capture multimodal interactions. In addition, unlike previous work that is either only pre-trained on image-text data or on fine-grained data with box-level annotations, we present a two-stage pre-training strategy that uses both these kinds of data efficiently: (i) coarse-grained pre-training based on image-text data; followed by (ii) fine-grained pre-training based on image-text-box data. We conduct comprehensive experiments on a wide range of VL tasks, ranging from VQA, image captioning, and retrieval, to phrase grounding, referring expression comprehension, and object detection. Using deep multimodal fusion coupled with the two-stage pre-training, FIBER provides consistent performance improvements over strong baselines across all tasks, often outperforming methods using magnitudes more data. Code is released at https://github.com/microsoft/FIBER.

In learning with recurrent or very deep feed-forward networks, employing unitary matrices in each layer can be very effective at maintaining long-range stability. However, restricting network parameters to be unitary typically comes at the cost of expensive parameterizations or increased training runtime. We propose instead an efficient method based on rank-$k$ updates -- or their rank-$k$ approximation -- that maintains performance at a nearly optimal training runtime. We introduce two variants of this method, named Direct (projUNN-D) and Tangent (projUNN-T) projected Unitary Neural Networks, that can parameterize full $N$-dimensional unitary or orthogonal matrices with a training runtime scaling as $O(kN^2)$. Our method either projects low-rank gradients onto the closest unitary matrix (projUNN-T) or transports unitary matrices in the direction of the low-rank gradient (projUNN-D). Even in the fastest setting ($k=1$), projUNN is able to train a model's unitary parameters to reach comparable performances against baseline implementations. In recurrent neural network settings, projUNN closely matches or exceeds benchmarked results from prior unitary neural networks. Finally, we preliminarily explore projUNN in training orthogonal convolutional neural networks, which are currently unable to outperform state of the art models but can potentially enhance stability and robustness at large depth.

Deep learning is increasingly moving towards a transfer learning paradigm whereby large foundation models are fine-tuned on downstream tasks, starting from an initialization learned on the source task. But an initialization contains relatively little information about the source task, and does not reflect the belief that our knowledge of the source task should affect the locations and shape of optima on the downstream task.Instead, we show that we can learn highly informative posteriors from the source task, through supervised or self-supervised approaches, which then serve as the basis for priors that modify the whole loss surface on the downstream task. This simple modular approach enables significant performance gains and more data-efficient learning on a variety of downstream classification and segmentation tasks, serving as a drop-in replacement for standard pre-training strategies. These highly informative priors also can be saved for future use, similar to pre-trained weights, and stand in contrast to the zero-mean isotropic uninformative priors that are typically used in Bayesian deep learning.

Deep learning (DL) and reinforcement learning (RL) methods seem to be a part of indispensable factors to achieve human-level or super-human AI systems. On the other hand, both DL and RL have strong connections with our brain functions and with neuroscientific findings. In this review, we summarize talks and discussions in the "Deep Learning and Reinforcement Learning" session of the symposium, International Symposium on Artificial Intelligence and Brain Science. In this session, we discussed whether we can achieve comprehensive understanding of human intelligence based on the recent advances of deep learning and reinforcement learning algorithms. Speakers contributed to provide talks about their recent studies that can be key technologies to achieve human-level intelligence.

Multi-modal reasoning systems rely on a pre-trained object detector to extract regions of interest from the image. However, this crucial module is typically used as a black box, trained independently of the downstream task and on a fixed vocabulary of objects and attributes. This makes it challenging for such systems to capture the long tail of visual concepts expressed in free form text. In this paper we propose MDETR, an end-to-end modulated detector that detects objects in an image conditioned on a raw text query, like a caption or a question. We use a transformer-based architecture to reason jointly over text and image by fusing the two modalities at an early stage of the model. We pre-train the network on 1.3M text-image pairs, mined from pre-existing multi-modal datasets having explicit alignment between phrases in text and objects in the image. We then fine-tune on several downstream tasks such as phrase grounding, referring expression comprehension and segmentation, achieving state-of-the-art results on popular benchmarks. We also investigate the utility of our model as an object detector on a given label set when fine-tuned in a few-shot setting. We show that our pre-training approach provides a way to handle the long tail of object categories which have very few labelled instances. Our approach can be easily extended for visual question answering, achieving competitive performance on GQA and CLEVR. The code and models are available at https://github.com/ashkamath/mdetr.

Self-supervised learning (SSL) is rapidly closing the gap with supervised methods on large computer vision benchmarks. A successful approach to SSL is to learn embeddings which are invariant to distortions of the input sample. However, a recurring issue with this approach is the existence of trivial constant solutions. Most current methods avoid such solutions by careful implementation details. We propose an objective function that naturally avoids collapse by measuring the cross-correlation matrix between the outputs of two identical networks fed with distorted versions of a sample, and making it as close to the identity matrix as possible. This causes the embedding vectors of distorted versions of a sample to be similar, while minimizing the redundancy between the components of these vectors. The method is called Barlow Twins, owing to neuroscientist H. Barlow’s redundancy-reduction principle applied to a pair of identical networks. Barlow Twins does not require large batches nor asymmetry between the network twins such as a predictor network, gradient stopping, or a moving average on the weight updates. Intriguingly it benefits from very high-dimensional output vectors. Barlow Twins outperforms previous methods on ImageNet for semi-supervised classification in the low-data regime, and is on par with current state of the art for ImageNet classification with a linear classifier head, and for transfer tasks of classification and object detection.

The task of image generation started receiving some attention from artists and designers, providing inspiration for new creations. However, exploiting the results of deep generative models such as Generative Adversarial Networks can be long and tedious given the lack of existing tools. In this work, we propose a simple strategy to inspire creators with new generations learned from a dataset of their choice, while providing some control over the output. We design a simple optimization method to find the optimal latent parameters corresponding to the closest generation to any input inspirational image. Specifically, we allow the generation given an inspirational image of the user's choosing by performing several optimization steps to recover optimal parameters from the model's latent space. We tested several exploration methods from classical gradient descents to gradient-free optimizers. Many gradient-free optimizers just need comparisons (better/worse than another image), so they can even be used without numerical criterion nor inspirational image, only with human preferences. Thus, by iterating on one's preferences we can make robust facial composite or fashion generation algorithms. Our results on four datasets of faces, fashion images, and textures show that satisfactory images are effectively retrieved in most cases.

An important component of autoencoder methods is the method by which the information capacity of the latent representation is minimized or limited. In this work, the rank of the covariance matrix of the codes is implicitly minimized by relying on the fact that gradient descent learning in multi-layer linear networks leads to minimum-rank solutions. By inserting a number of extra linear layers between the encoder and the decoder, the system spontaneously learns representations with a low effective dimension. The model, dubbed Implicit Rank-Minimizing Autoencoder (IRMAE), is simple, deterministic, and learns continuous latent space. We demonstrate the validity of the method on several image generation and representation learning tasks.

Learning with complete or partial supervision is power- ful but relies on ever-growing human annotation efforts. As a way to mitigate this serious problem, as well as to serve specific applications, unsupervised learning has emerged as an important field of research. In computer vision, unsu- pervised learning comes in various guises. We focus here on the unsupervised discovery and matching of object cate- gories among images in a collection, following the work of Cho et al. [12]. We show that the original approach can be reformulated and solved as a proper optimization problem. Experiments on several benchmarks establish the merit of our approach.

In this paper we discuss several forms of spatiotemporal convolutions for video analysis and study their effects on action recognition. Our motivation stems from the observation that 2D CNNs applied to individual frames of the video have remained solid performers in action recognition. In this work we empirically demonstrate the accuracy advantages of 3D CNNs over 2D CNNs within the framework of residual learning. Furthermore, we show that factorizing the 3D convolutional filters into separate spatial and temporal components yields significantly gains in accuracy. Our empirical study leads to the design of a new spatiotemporal convolutional block "R(2+1)D" which produces CNNs that achieve results comparable or superior to the state-of-the-art on Sports-1M, Kinetics, UCF101, and HMDB51.

We analyze numerically the training dynamics of deep neural networks (DNN) by using methods developed in statistical physics of glassy systems. The two main issues we address are the complexity of the loss-landscape and of the dynamics within it, and to what extent DNNs share similarities with glassy systems. Our findings, obtained for different architectures and data-sets, suggest that during the training process the dynamics slows down because of an increasingly large number of flat directions. At large times, when the loss is approaching zero, the system diffuses at the bottom of the landscape. Despite some similarities with the dynamics of mean-field glassy systems, in particular, the absence of barrier crossing, we find distinctive dynamical behaviors in the two cases, thus showing that the statistical properties of the corresponding loss and energy landscapes are different. In contrast, when the network is under-parametrized we observe a typical glassy behavior, thus suggesting the existence of different phases depending on whether the network is under-parametrized or over-parametrized.

Deep latent variable models, trained using variational autoencoders or generative adversarial networks, are now a key technique for representation learning of continuous structures. However, applying similar methods to discrete structures, such as text sequences or discretized images, has proven to be more challenging. In this work, we propose a more flexible method for training deep latent variable models of discrete structures. Our approach is based on the recently proposed Wasserstein Autoencoder (WAE) which formalizes adversarial autoencoders as an optimal transport problem. We first extend this framework to model discrete sequences, and then further explore different learned priors targeting a controllable representation. Unlike many other latent variable generative models for text, this adversarially regularized autoencoder (ARAE) allows us to generate fluent textual outputs as well as perform manipulations in the latent space to induce change in the output space. Finally we show that the latent representation can be trained to perform unaligned textual style transfer, giving improvements both in automatic measures and human evaluation.

Modern deep transfer learning approaches have mainly focused on learning generic feature vectors from one task that are transferable to other tasks, such as word embeddings in language and pretrained convolutional features in vision. However, these approaches usually transfer unary features and largely ignore more structured graphical representations. This work explores the possibility of learning generic latent relational graphs that capture dependencies between pairs of data units (e.g., words or pixels) from large-scale unlabeled data and transferring the graphs to downstream tasks. Our proposed transfer learning framework improves performance on various tasks including question answering, natural language inference, sentiment analysis, and image classification. We also show that the learned graphs are generic enough to be transferred to different embeddings on which the graphs have not been trained (including GloVe embeddings, ELMo embeddings, and task-specific RNN hidden units), or embedding-free units such as image pixels.

This paper shows that simply prescribing "none of the above" labels to unlabeled data has a beneficial regularization effect to supervised learning. We call it universum prescription by the fact that the prescribed labels cannot be one of the supervised labels. In spite of its simplicity, universum prescription obtained competitive results in training deep convolutional networks for CIFAR-10, CIFAR-100, STL-10 and ImageNet datasets. A qualitative justification of these approaches using Rademacher complexity is presented. The effect of a regularization parameter — probability of sampling from unlabeled data — is also studied empirically.

The dominant approach for many NLP tasks are recurrent neural networks, in particular LSTMs, and convolutional neural networks. However, these architectures are rather shallow in comparison to the deep convolutional networks which have pushed the state-of-the-art in computer vision. We present a new architecture (VDCNN) for text processing which operates directly at the character level and uses only small convolutions and pooling operations. We are able to show that the performance of this model increases with depth: using up to 29 convolutional layers, we report improvements over the state-of-the-art on several public text classification tasks. To the best of our knowledge, this is the first time that very deep convolutional nets have been applied to text processing.

The ability to predict and therefore to anticipate the future is an important attribute of intelligence. It is also of utmost importance in real-time systems, e.g. in robotics or autonomous driving, which depend on visual scene understanding for decision making. While prediction of the raw RGB pixel values in future video frames has been studied in previous work, here we introduce the novel task of predicting semantic segmentations of future frames. Given a sequence of video frames, our goal is to predict segmentation maps of not yet observed video frames that lie up to a second or further in the future. We develop an autoregressive convolutional neural network that learns to iteratively generate multiple frames. Our results on the Cityscapes dataset show that directly predicting future segmentations is substantially better than predicting and then segmenting future RGB frames. Prediction results up to half a second in the future are visually convincing and are much more accurate than those of a baseline based on warping semantic segmentations using optical flow.

Using unitary (instead of general) matrices in artificial neural networks (ANNs) is a promising way to solve the gradient explosion/vanishing problem, as well as to enable ANNs to learn long-term correlations in the data. This approach appears particularly promising for Recurrent Neural Networks (RNNs). In this work, we present a new architecture for implementing an Efficient Unitary Neural Network (EUNNs); its main advantages can be summarized as follows. Firstly, the representation capacity of the unitary space in an EUNN is fully tunable, ranging from a subspace of SU(N) to the entire unitary space. Secondly, the computational complexity for training an EUNN is merely $\mathcal{O}(1)$ per parameter. Finally, we test the performance of EUNNs on the standard copying task, the pixel-permuted MNIST digit recognition benchmark as well as the Speech Prediction Test (TIMIT). We find that our architecture significantly outperforms both other state-of-the-art unitary RNNs and the LSTM architecture, in terms of the final performance and/or the wall-clock training speed. EUNNs are thus promising alternatives to RNNs and LSTMs for a wide variety of applications.

Convolutional neural networks (CNNs) are a standard component of many current state-of-the-art Large Vocabulary Continuous Speech Recognition (LVCSR) systems. However, CNNs in LVCSR have not kept pace with recent advances in other domains where deeper neural networks provide superior performance. In this paper we propose a number of architectural advances in CNNs for LVCSR. First, we introduce a very deep convolutional network architecture with up to 14 weight layers. There are multiple convolutional layers before each pooling layer, with small 3×3 kernels, inspired by the VGG Imagenet 2014 architecture. Then, we introduce multilingual CNNs with multiple untied layers. Finally, we introduce multi-scale input features aimed at exploiting more context at negligible computational cost. We evaluate the improvements first on a Babel task for low resource speech recognition, obtaining an absolute 5.77% WER improvement over the baseline PLP DNN by training our CNN on the combined data of six different languages. We then evaluate the very deep CNNs on the Hub5'00 benchmark (using the 262 hours of SWB-1 training data) achieving a word error rate of 11.8% after cross-entropy training, a 1.4% WER improvement (10.6% relative) over the best published CNN result so far.

We consider the hashing mechanism for constructing binary embeddings, that involves pseudo-random projections followed by nonlinear (sign function) mappings. The pseudo-random projection is described by a matrix, where not all entries are independent random variables but instead a fixed “budget of randomness” is distributed across the matrix. Such matrices can be edfficiently stored in sub-quadratic or even linear space, provide reduction in randomness usage (i.e. number of required random values), and very often lead to computational speed ups. We prove several theoretical results showing that projections via various structured matrices followed by nonlinear mappings accurately preserve the angular distance between input high-dimensional vectors. To the best of our knowledge, these results are the first that give theoretical ground for the use of general structured matrices in the nonlinear setting. We empirically verify our theoretical findings and show the dependence of learning via structured hashed projections on the performance of neural network as well as nearest neighbor classifier.

Although RNNs have been shown to be power- ful tools for processing sequential data, finding architectures or optimization strategies that al- low them to model very long term dependencies is still an active area of research. In this work, we carefully analyze two synthetic datasets orig- inally outlined in (Hochreiter & Schmidhuber, 1997) which are used to evaluate the ability of RNNs to store information over many time steps. We explicitly construct RNN solutions to these problems, and using these constructions, illumi- nate both the problems themselves and the way in which RNNs store different types of information in their hidden states. These constructions fur- thermore explain the success of recent methods that specify unitary initializations or constraints on the transition matrices.

We propose a deep generative model for learning to distill the hidden factors of variation within a set of labeled observations into two complementary codes. One code describes the factors of variation relevant to solving a specified task. The other code describes the remaining factors of variation that are irrelevant to solving this task. The only available source of supervision during the training process comes from our ability to distinguish among different observations belonging to the same category. Concrete examples include multiple images of the same object from different viewpoints, or multiple speech samples from the same speaker. In both of these instances, the factors of variation irrelevant to classification are implicitly expressed by intra-class variabilities, such as the relative position of an object in an image, or the linguistic content of an utterance. Most existing approaches for solving this problem rely heavily on having access to pairs of observations only sharing a single factor of variation, e.g. different objects observed in the exact same conditions. This assumption is often not encountered in realistic settings where data acquisition is not controlled and labels for the uninformative components are not available. In this work, we propose to overcome this limitation by augmenting deep convolutional autoencoders with a form of adversarial training. Both factors of variation are implicitly captured in the organization of the learned embedding space, and can be used for solving single-image analogies. Experimental results on synthetic and real datasets show that the proposed method is capable of disentangling the influences of style and content factors using a flexible representation, as well as generalizing to unseen styles or content classes.

We present a method for extracting depth information from a rectified image pair. Our approach focuses on the first stage of many stereo algorithms: the matching cost computation. We approach the problem by learning a similarity measure on small image patches using a convolutional neural network. Training is carried out in a supervised manner by constructing a binary classification data set with examples of similar and dissimilar pairs of patches. We examine two network architectures for this task: one tuned for speed, the other for accuracy. The output of the convolutional neural network is used to initialize the stereo matching cost. A series of post-processing steps follow: cross-based cost aggregation, semiglobal matching, a left-right consistency check, subpixel enhancement, a median filter, and a bilateral filter. We evaluate our method on the KITTI 2012, KITTI 2015, and Middlebury stereo data sets and show that it outperforms other approaches on all three data sets.

A complex-valued convolutional network (convnet) implements the repeated application of the following composition of three operations, recursively applying the composition to an input vector of nonnegative real numbers: (1) convolution with complex-valued vectors, followed by (2) taking the absolute value of every entry of the resulting vectors, followed by (3) local averaging. For processing real-valued random vectors, complex-valued convnets can be viewed as data-driven multiscale windowed power spectra, data-driven multiscale windowed absolute spectra, data-driven multiwavelet absolute values, or (in their most general configuration) data-driven nonlinear multiwavelet packets. Indeed, complex-valued convnets can calculate multiscale windowed spectra when the convnet filters are windowed complex-valued exponentials. Standard real-valued convnets, using rectified linear units (ReLUs), sigmoidal (e.g., logistic or tanh) nonlinearities, or max pooling, for example, do not obviously exhibit the same exact correspondence with data-driven wavelets (whereas for complex-valued convnets, the correspondence is much more than just a vague analogy). Courtesy of the exact correspondence, the remarkably rich and rigorous body of mathematical analysis for wavelets applies directly to (complex-valued) convnets.

We study the connection between the highly non-convex loss function of a simple model of the fully-connected feed-forward neural network and the Hamiltonian of the spherical spin-glass model under the assumptions of: i) variable independence, ii) redundancy in network parametrization, and iii) uniformity. These assumptions enable us to explain the complexity of the fully decoupled neural network through the prism of the results from random matrix theory. We show that for large-size decoupled networks the lowest critical values of the random loss function form a layered structure and they are located in a well-defined band lower-bounded by the global minimum. The number of local minima outside that band diminishes exponentially with the size of the network. We empirically verify that the mathematical model exhibits similar behavior as the computer simulations, despite the presence of high dependencies in real networks. We conjecture that both simulated annealing and SGD converge to the band of low critical points, and that all critical points found there are local minima of high quality measured by the test error. This emphasizes a major difference between large- and small-size networks where for the latter poor quality local minima have non-zero probability of being recovered. Finally, we prove that recovering the global minimum becomes harder as the network size increases and that it is in practice irrelevant as global minimum often leads to overfitting.

Deep learning has enjoyed a resurgence of interest in the last few years for such applications as image and speech recognition, or natural language processing. The vast majority of practical applications of deep learning focus on supervised learning, where the supervised loss function is minimized using stochastic gradient descent. The properties of this highly non-convex loss function, such as its landscape and the behavior of critical points (maxima, minima, and saddle points), as well as the reason why large- and small-size networks achieve radically different practical performance, are however very poorly understood. It was only recently shown that new results in spin-glass theory potentially may provide an explanation for these problems by establishing a connection between the loss function of the neural networks and the Hamiltonian of the spherical spin-glass models. The connection between both models relies on a number of possibly unrealistic assumptions, yet the empirical evidence suggests that the connection may exist in real. The question we pose is whether it is possible to drop some of these assumptions to establish a stronger connection between both models.

Recent state-of-the-art performance on human-body pose estimation has been achieved with Deep Convolutional Networks (ConvNets). Traditional ConvNet architectures include pooling and sub-sampling layers which reduce computational requirements, introduce invariance and prevent over-training. These benefits of pooling come at the cost of reduced localization accuracy. We introduce a novel architecture which includes an efficient ‘position refinement’ model that is trained to estimate the joint offset location within a small region of the image. This refinement model is jointly trained in cascade with a state-of-the-art ConvNet model [21] to achieve improved accuracy in human joint location estimation. We show that the variance of our detector approaches the variance of human annotations on the FLIC [20] dataset and outperforms all existing approaches on the MPII-human-pose dataset [1].

We present a method for extracting depth information from a rectified image pair. We train a convolutional neural network to predict how well two image patches match and use it to compute the stereo matching cost. The cost is refined by cross-based cost aggregation and semiglobal matching, followed by a left-right consistency check to eliminate errors in the occluded regions. Our stereo method achieves an error rate of 2.61% on the KITTI stereo dataset and is currently (August 2014) the top performing method on this dataset.

This paper presents a single-channel source separation method that extends the ideas of Nonnegative Matrix Factorization (NMF). We interpret the approach of audio demixing via NMF as a cascade of a pooled analysis operator, given for example by the magnitude spectrogram, and a synthesis operators given by the matrix decomposition. Instead of imposing the temporal consistency of the decomposition through sophisticated structured penalties in the synthesis stage, we propose to change the analysis operator for a deep scattering representation, where signals are represented at several time resolutions. This new signal representation is invariant to smooth changes in the signal, consistent with its temporal dynamics. We evaluate the proposed approach in a speech separation task obtaining promising results.

Current state-of-the-art classification and detection algorithms train deep convolutional networks using labeled data. In this work we study unsupervised feature learning with convolutional networks in the context of temporally coherent unlabeled data. We focus on feature learning from unlabeled video data, using the assumption that adjacent video frames contain semantically similar information. This assumption is exploited to train a convolutional pooling auto-encoder regularized by slowness and sparsity priors. We establish a connection between slow feature learning and metric learning. Using this connection we define "temporal coherence" -- a criterion which can be used to set hyper-parameters in a principled and automated manner. In a transfer learning experiment, we show that the resulting encoder can be used to define a more semantically coherent metric without the use of labels.

Training deep feature hierarchies to solve supervised learning tasks has achieving state of the art performance on many problems in computer vision. However, a principled way in which to train such hierarchies in the unsupervised setting has remained elusive. In this work we suggest a new architecture and loss for training deep feature hierarchies that linearize the transformations observed in unlabelednatural video sequences. This is done by training a generative model to predict video frames. We also address the problem of inherent uncertainty in prediction by introducing a latent variables that are non-deterministic functions of the input into the network architecture.

We study the problem of stochastic optimization for deep learning in the parallel computing environment under communication constraints. A new algorithm is proposed in this setting where the communication and coordination of work among concurrent processes (local workers), is based on an elastic force which links the parameters they compute with a center variable stored by the parameter server (master). The algorithm enables the local workers to perform more exploration, i.e. the algorithm allows the local variables to fluctuate further from the center variable by reducing the amount of communication between local workers and the master. We empirically demonstrate that in the deep learning setting, due to the existence of many local optima, allowing more exploration can lead to the improved performance. We propose synchronous and asynchronous variants of the new algorithm. We provide the stability analysis of the asynchronous variant in the round-robin scheme and compare it with the more common parallelized method ADMM. We show that the stability of EASGD is guaranteed when a simple stability condition is satisfied, which is not the case for ADMM. We additionally propose the momentum-based version of our algorithm that can be applied in both synchronous and asynchronous settings. Asynchronous variant of the algorithm is applied to train convolutional neural networks for image classification on the CIFAR and ImageNet datasets. Experiments demonstrate that the new algorithm accelerates the training of deep architectures compared to DOWNPOUR and other common baseline approaches and furthermore is very communication efficient.

This article offers an empirical exploration on the use of character-level convolutional networks (ConvNets) for text classification. We constructed several large-scale datasets to show that character-level convolutional networks could achieve state-of-the-art or competitive results. Comparisons are offered against traditional models such as bag of words, n-grams and their TFIDF variants, and deep learning models such as word-based ConvNets and recurrent neural networks.

Pooling operators construct non-linear representations by cascading a redundant linear transform, followed by a point-wise nonlinearity and a local aggregation, typically implemented with a \ell_p norm. Their efficiency in recognition architectures is based on their ability to locally contract the input space, but also on their capacity to retain as much stable information as possible. We address this latter question by computing the upper and lower Lipschitz bounds of \ell_p pooling operators for p=1, 2, ∞as well as their half-rectified equivalents, which give sufficient conditions for the design of invertible pooling layers. Numerical experiments on MNIST and image patches confirm that pooling layers can be inverted with phase recovery algorithms. Moreover, the regularity of the inverse pooling, controlled by the lower Lipschitz constant, is empirically verified with a nearest neighbor regression.

We present a novel method for real-time continuous pose recovery of markerless complex articulable objects from a single depth image. Our method consists of the following stages: a randomized decision forest classifier for image segmentation, a robust method for labeled dataset generation, a convolutional network for dense feature extraction, and finally an inverse kinematics stage for stable real-time pose recovery. As one possible application of this pipeline, we show state-of-the-art results for real-time puppeteering of a skinned hand-model.

Pedestrian detection is a problem of considerable practical interest. Adding to the list of successful applications of deep learning methods to vision, we report state-of-the-art and competitive results on all major pedestrian datasets with a convolutional network model. The model uses a few new twists, such as multi-stage features, connections that skip layers to integrate global shape information with local distinctive motif information, and an unsupervised method based on convolutional sparse coding to pre-train the filters at each stage.

Among the different methods producing superpixel segmentations of an image, the graph-based approach of Felzenszwalb and Huttenlocher is broadly employed. One of its interesting properties is that the regions are computed in a greedy manner in quasi-linear time by using a minimum spanning tree. The algorithm may be trivially extended to video segmentation by considering a video as a 3D volume, however, this can not be the case for causal segmentation, when subsequent frames are unknown. In a framework exploiting minimum spanning trees all along, we propose an efficient video segmentation approach that computes temporally consistent pixels in a causal manner, filling the need for causal and real time applications.

The performance of stochastic gradient descent (SGD) depends critically on how learning rates are tuned and decreased over time. We propose a method to automatically adjust multiple learning rates so as to minimize the expected error at any one time. The method relies on local gradient variations across samples. In our approach, learning rates can increase as well as decrease, making it suitable for non-stationary problems. Using a number of convex and non-convex learning tasks, we show that the resulting algorithm matches the performance of the best settings obtained through systematic search, and effectively removes the need for learning rate tuning.

We introduce DropConnect, a generalization of DropOut, for regularizing large fully-connected layers within neural networks. When training with Dropout, a randomly selected subset of activations are set to zero within each layer. DropConnect instead sets a randomly selected subset of weights within the network to zero. Each unit thus receives input from a random subset of units in the previous layer. We derive a bound on the generalization performance of both Dropout and DropConnect. We then evaluate DropConnect on a range of datasets, comparing to Dropout, and show state-of-the-art results on several image recoginition benchmarks can be obtained by aggregating multiple DropConnect-trained models.

Scene labeling consists of labeling each pixel in an image with the category of the object it belongs to. We propose a method that uses a multiscale convolutional network trained from raw pixels to extract dense feature vectors that encode regions of multiple sizes centered on each pixel. The method alleviates the need for engineered features, and produces a powerful representation that captures texture, shape, and contextual information. We report results using multiple postprocessing methods to produce the final labeling. Among those, we propose a technique to automatically retrieve, from a pool of segmentation components, an optimal set of components that best explain the scene; these components are arbitrary, for example, they can be taken from a segmentation tree or from any family of oversegmentations. The system yields record accuracies on the SIFT Flow dataset (33 classes) and the Barcelona dataset (170 classes) and near-record accuracy on Stanford background dataset (eight classes), while being an order of magnitude faster than competing approaches, producing a $(320\times 240)$ image labeling in less than a second, including feature extraction.

Invariant representations in object recognition systems are generally obtained by pooling feature vectors over spatially local neighborhoods. But pooling is not local in the feature vector space, so that widely dissimilar features may be pooled together if they are in nearby locations. Recent approaches rely on sophisticated encoding methods and more specialized codebooks (or dictionaries), e.g., learned on subsets of descriptors which are close in feature space, to circumvent this problem. In this work, we argue that a common trait found in much recent work in image recognition or retrieval is that it leverages locality in feature space on top of purely spatial locality. We propose to apply this idea in its simplest form to an object recognition system based on the spatial pyramid framework, to increase the performance of small dictionaries with very little added engineering. State-of-the-art results on several object recognition benchmarks show the promise of this approach.

We apply Convolutional Networks (ConvNets) to the task of traffic sign classification as part of the GTSRB competition. ConvNets are biologically-inspired multi-stage architectures that automatically learn hierarchies of invariant features. While many popular vision approaches use hand-crafted features such as HOG or SIFT, ConvNets learn features at every level from data that are tuned to the task at hand. The traditional ConvNet architecture was modified by feeding 1ststage features in addition to 2ndstage features to the classifier. The system yielded the 2nd-best accuracy of 98.97% during phase I of the competition (the best entry obtained 98.98%), above the human performance of 98.81%, using 32×32 color input images. Experiments conducted after phase 1 produced a new record of 99.17% by increasing the network capacity, and by using greyscale images instead of color. Interestingly, random features still yielded competitive results (97.33%).

This work describes a conceptually simple method for structured sparse coding and dictionary design. Supposing a dictionary with K atoms, we introduce a structure as a set of penalties or interactions between every pair of atoms. We describe modifications of standard sparse coding algorithms for inference in this setting, and describe experiments showing that these algorithms are efficient. We show that interesting dictionaries can be learned for interactions that encode tree structures or locally connected structures. Finally, we show that our framework allows us to learn the values of the interactions from the data, rather than having them pre-specified.

Many successful models for scene or object recognition transform low-level descriptors (such as Gabor filter responses, or SIFT descriptors) into richer representations of intermediate complexity. This process can often be broken down into two steps: (1) a coding step, which performs a pointwise transformation of the descriptors into a representation better adapted to the task, and (2) a pooling step, which summarizes the coded features over larger neighborhoods. Several combinations of coding and pooling schemes have been proposed in the literature. The goal of this paper is threefold. We seek to establish the relative importance of each step of mid-level feature extraction through a comprehensive cross evaluation of several types of coding modules (hard and soft vector quantization, sparse coding) and pooling schemes (by taking the average, or the maximum), which obtains state-of-the-art performance or better on several recognition benchmarks. We show how to improve the best performing coding scheme by learning a supervised discriminative dictionary for sparse coding. We provide theoretical and empirical insight into the remarkable performance of max pooling. By teasing apart components shared by modern mid-level feature extractors, our approach aims to facilitate the design of better recognition architectures.

We present a novel “hybrid Hessian” six-degrees-of-freedom simultaneous localization and mapping (SLAM) algorithm. Our method allows for the smooth trade-off of accuracy for efficiency and for the incorporation of GPS measurements during real-time operation, thereby offering significant advantages over other SLAM solvers. Like other stochastic SLAM methods, such as SGD and TORO, our technique is robust to local minima and eliminates the need for costly relinearizations of the map. Unlike other stochastic methods, but similar to exact solvers, such as iSAM, our technique is able to process position-only constraints, such as GPS measurements, without introducing systematic distortions in the map. We present results from the Google Street View database, and compare our method with results from TORO. We show that our solver is able to achieve higher accuracy while operating within real-time bounds. In addition, as far as we are aware, this is the first stochastic SLAM solver capable of processing GPS constraints in real-time.

In this paper we present a scalable hardware architecture to implement large-scale convolutional neural networks and state-of-the-art multi-layered artificial vision systems. This system is fully digital and is a modular vision engine with the goal of performing real-time detection, recognition and segmentation of mega-pixel images. We present a performance comparison between a software, FPGA and ASIC implementation that shows a speed up in custom hardware implementations.

Intelligent tasks, such as visual perception, auditory perception, and language understanding require the construction of good internal representations of the world (or "features")? which must be invariant to irrelevant variations of the input while, preserving relevant information. A major question for Machine Learning is how to learn such good features automatically. Convolutional Networks (ConvNets) are a biologically-inspired trainable architecture that can learn invariant features. Each stage in a ConvNets is composed of a filter bank, some nonlinearities, and feature pooling layers. With multiple stages, a ConvNet can learn multi-level hierarchies of features. While ConvNets have been successfully deployed in many commercial applications from OCR to video surveillance, they require large amounts of labeled training samples. We describe new unsupervised learning algorithms, and new non-linear stages that allow ConvNets to be trained with very few labeled samples. Applications to visual object recognition and vision navigation for off-road mobile robots are described.

We propose an unsupervised method for learning multi-stage hierarchies of sparse convolutional features. While sparse coding has become an increasingly popular method for learning visual features, it is most often trained at the patch level. Applying the resulting filters convolutionally results in highly redundant codes because overlapping patches are encoded in isolation. By training convolutionally over large image windows, our method reduces the redudancy between feature vectors at neighboring locations and improves the efficiency of the overall representation. In addition to a linear decoder that reconstructs the image from sparse features, our method trains an efficient feed-forward encoder that predicts quasi-sparse features from the input. While patch-based training rarely produces anything but oriented edge detectors, we show that convolutional training produces highly diverse filters, including center-surround filters, corner detectors, cross detectors, and oriented grating detectors. We show that using these filters in multi-stage convolutional network architecture improves performance on a number of visual recognition and detection tasks.

Score Matching is a recently-proposed criterion for training high-dimensional density models for which maximum likelihood training is intractable. It has been applied to learning natural image statistics but has so-far been limited to simple models due to the difficulty of differentiating the loss with respect to the model parameters. We show how this differentiation can be automated with an extended version of the double-backpropagation algorithm. In addition, we introduce a regularization term for the Score Matching loss that enables its use for a broader range of problem by suppressing instabilities that occur with finite training sample sizes and quantized input values. Results are reported for image denoising and super-resolution.

Several recently-proposed architectures for high-performance object recognition are composed of two main stages: a feature extraction stage that extracts locally-invariant feature vectors from regularly spaced image patches, and a somewhat generic supervised classifier. The first stage is often composed of three main modules: (1) a bank of filters (often oriented edge detectors); (2) a non-linear transform, such as a point-wise squashing functions, quantization, or normalization; (3) a spatial pooling operation which combines the outputs of similar filters over neighboring regions. We propose a method that automatically learns such feature extractors in an unsupervised fashion by simultaneously learning the filters and the pooling units that combine multiple filter outputs together. The method automatically generates topographic maps of similar filters that extract features of orientations, scales, and positions. These similar filters are pooled together, producing locally-invariant outputs. The learned feature descriptors give comparable results as SIFT on image recognition tasks for which SIFT is well suited, and better results than SIFT on tasks for which SIFT is less well suited.

Convolutional networks (ConvNets) are biologically inspired hierarchical architectures that can be trained to perform a variety of detection, recognition and segmentation tasks. ConvNets have a feed-forward architecture consisting of multiple linear convolution filters interspersed with pointwise non-linear squashing functions. This paper presents an efficient implementation of ConvNets on a low-end DSP-oriented field programmable gate array (FPGA). The implementation exploits the inherent parallelism of ConvNets and takes full advantage of multiple hardware multiply accumulate units on the FPGA. The entire system uses a single FPGA with an external memory module, and no extra parts. A network compiler software was implemented, which takes a description of a trained ConvNet and compiles it into a sequence of instructions for the ConvNet Processor (CNP). A ConvNet face detection system was implemented and tested. Face detection on a 512 times 384 frame takes 100 ms (10 frames per second), which corresponds to an average performance of 3.4 times 109connections per second for this 340 million connection network. The design can be used for low-power, lightweight embedded vision systems for micro-UAVs and other small robots.

In many recent object recognition systems, feature extraction stages are generally composed of a filter bank, a non-linear transformation, and some sort of feature pooling layer. Most systems use only one stage of feature extraction in which the filters are hard-wired, or two stages where the filters in one or both stages are learned in supervised or unsupervised mode. This paper addresses three questions: 1. How does the non-linearities that follow the filter banks influence the recognition accuracy? 2. does learning the filter banks in an unsupervised or supervised manner improve the performance over random filters or hardwired filters? 3. Is there any advantage to using an architecture with two stages of feature extraction, rather than one? We show that using non-linearities that include rectification and local contrast normalization is the single most important ingredient for good accuracy on object recognition benchmarks. We show that two stages of feature extraction yield better accuracy than one. Most surprisingly, we show that a two-stage system with random filters can yield almost 63% recognition rate on Caltech-101, provided that the proper non-linearities and pooling layers are used. Finally, we show that with supervised refinement, the system achieves state-of-the-art performance on NORB dataset (5.6%) and unsupervised pre-training followed by supervised refinement produces good accuracy on Caltech-101 (> 65%), and the lowest known error rate on the undistorted, unprocessed MNIST dataset (0.53%).

No abstract available.

We describe an efficient, low-cost, low-overhead system for robot localization in complex visual environments. Our system augments wheel odometry with visual orientation tracking to yield localization accuracy comparable with ldquopurerdquo visual odometry at a fraction of the cost. Such a system is well-suited to consumer-level robots, small form-factor robots, extraterrestrial rovers, and other platforms with limited computational resources. Our system also benefits high-end multiprocessor robots by leaving ample processor time on all camera-computer pairs to perform other critical visual tasks, such as obstacle detection. Experimental results are shown for outdoor, off-road loops on the order of 200 meters. Comparisons are made with corresponding results from a state-of-the-art pure visual odometer.

Energy-based learning (EBL) is a general framework to describe supervised and unsupervised training methods for probabilistic and non-probabilistic factor graphs. An energy-based model associates a scalar energy to configurations of inputs, outputs, and latent variables. Learning machines can be constructed by assembling modules and loss functions. Gradient-based learning procedures are easily implemented through semi-automatic differentiation of complex models constructed by assembling predefined modules. We introduce an open-source and cross-platform C++ library called EBLearn to enable the construction of energy-based learning models. EBLearn is composed of two major components, libidx: an efficient and flexible multi-dimensional tensor library, and libeblearn: an object-oriented library of trainable modules and learning algorithms. The latter has facilities for such models as convolutional networks, as well as for image processing. It also provides graphical display functions.

We present a learning-based approach for long-range vision that is able to accurately classify complex terrain at distances up to the horizon, thus allowing high-level strategic planning. A deep belief network is trained with unsupervised data and a reconstruction criterion to extract features from an input image, and the features are used to train a realtime classifier to predict traversability. The online supervision is given by a stereo module that provides robust labels for nearby areas up to 12 meters distant. The approach was developed and tested on the LAGR mobile robot.

Recent advances in self-supervised learning have enabled very long-range visual detection of obstacles and pathways (to 100 meters or more). Unfortunately, the category and range of regions at such large distances come with a considerable amount of uncertainty. We present a mapping and planning system that accurately represents range and category uncertainties, and accumulates the evidence from multiple frames in a principled way. The system relies on a hyperbolicpolar map centered on the robot with a 200 m radius. Map cells are histograms that accumulate evidence obtained from a self-supervised object classifier operating on image windows. The performance of the system is demonstrated on the LAGR off-road robot platform.

We present an unsupervised method for learning a hierarchy of sparse feature detectors that are invariant to small shifts and distortions. The resulting feature extractor consists of multiple convolution filters, followed by a feature-pooling layer that computes the max of each filter output within adjacent windows, and a point-wise sigmoid non-linearity. A second level of larger and more invariant features is obtained by training the same algorithm on patches of features from the first level. Training a supervised classifier on these features yields 0.64% error on MNIST, and 54% average recognition rate on Caltech 101 with 30 training samples per category. While the resulting architecture is similar to convolutional networks, the layer-wise unsupervised training procedure alleviates the over-parameterization problems that plague purely supervised learning procedures, and yields good performance with very few labeled training samples.

The machine learning and pattern recognition communities are facing two challenges: solving the normalization problem, and solving the deep learning problem. The normalization problem is related to the difficulty of training probabilistic models over large spaces while keeping them properly normalized. In recent years, the ML and natural language communities have devoted considerable efforts to circumventing this problem by developing "un-normalized" learning models for tasks in which the output is highly structured (e.g. English sentences). This class of models was in fact originally developed during the 90's in the handwriting recognition community, and includes graph transformer networks, conditional random fields, hidden Markov SVMs, and maximum margin Markov networks. We describe these models within the unifying framework of "energy-based models" (EBM). The deep learning problem is related to the issue of training all the levels of a recognition system (e.g. segmentation, feature extraction, recognition, etc) in an integrated fashion. We first consider " traditional" methods for deep learning, such as convolutional networks and back-propagation, and show that, although they produce very low error rates for handwriting and object recognition, they require many training samples. We show that using unsupervised learning to initialize the layers of a deep network dramatically reduces the required number of training samples, particularly for such tasks as the recognition of everyday objects at the category level.

We describe an unsupervised learning algorithm for extracting sparse and locally shift-invariant features. We also devise a principled procedure for learning hierarchies of invariant features. Each feature detector is composed of a set of trainable convolutional filters followed by a max-pooling layer over non-overlapping windows, and a point-wise sigmoid non-linearity. A second stage of more invariant features is fed with patches provided by the first stage feature extractor, and is trained in the same way. The method is used to pre-train the first four layers of a deep convolutional network which achieves state-of-the-art performance on the MNIST dataset of handwritten digits. The final testing error rate is equal to 0.42%. Preliminary experiments on compression of bitonal document images show very promising results in terms of compression ratio and reconstruction error.

A novel probabilistic online learning framework for autonomous off-road robot navigation is proposed. The system is purely vision-based and is particularly designed for predicting traversability in unknown or rapidly changing environments. It uses self-supervised learning to quickly adapt to novel terrains after processing a small number of frames, and it can recognize terrain elements such as paths, man-made structures, and natural obstacles at ranges up to 30 meters. The system is developed on the LAGR mobile robot platform and the performance is evaluated using multiple metrics, including ground truth.

In many regression problems, the variable to be predicted depends not only on a sample-specific feature vector, but also on an unknown (latent) manifold that must satisfy known constraints. An example is house prices, which depend on the characteristics of the house, and on the desirability of the neighborhood, which is not directly measurable. The proposed method comprises two trainable components. The first one is a parametric model that predicts the "intrinsic" price of the house from its description. The second one is a smooth, non-parametric model of the latent "desirability" manifold. The predicted price of a house is the product of its intrinsic price and desirability. The two components are trained simultaneously using a deterministic form of the EM algorithm. The model was trained on a large dataset of houses from Los Angeles county. It produces better predictions than pure parametric and non-parametric models. It also produces useful estimates of the desirability surface at each location.

Unsupervised learning algorithms aim to discover the structure hidden in the data, and to learn representations that are more suitable as input to a supervised machine than the raw input. Many unsupervised methods are based on reconstructing the input from the representation, while constraining the representation to have certain desirable properties (e.g. low dimension, sparsity, etc). Others are based on approximating density by stochastically reconstructing the input from the representation. We describe a novel and efficient algorithm to learn sparse representations, and compare it theoretically and experimentally with a similar machines trained probabilistically, namely a Restricted Boltzmann Machine. We propose a simple criterion to compare and select different unsupervised machines based on the trade-off between the reconstruction error and the information content of the representation. We demonstrate this method by extracting features from a dataset of handwritten numerals, and from a dataset of natural image patches. We show that by stacking multiple levels of such machines and by training sequentially, high-order dependencies between the input variables can be captured.

Dimensionality reduction involves mapping a set of high dimensional input points onto a low dimensional manifold so that 'similar" points in input space are mapped to nearby points on the manifold. We present a method - called Dimensionality Reduction by Learning an Invariant Mapping (DrLIM) - for learning a globally coherent nonlinear function that maps the data evenly to the output manifold. The learning relies solely on neighborhood relationships and does not require any distancemeasure in the input space. The method can learn mappings that are invariant to certain transformations of the inputs, as is demonstrated with a number of experiments. Comparisons are made to other techniques, in particular LLE.

The detection and recognition of generic object categories with invariance to viewpoint, illumination, and clutter requires the combination of a feature extractor and a classifier. We show that architectures such as convolutional networks are good at learning invariant features, but not always optimal for classification, while Support Vector Machines are good at producing decision surfaces from wellbehaved feature vectors, but cannot learn complicated invariances. We present a hybrid system where a convolutional network is trained to detect and recognize generic objects, and a Gaussian-kernel SVM is trained from the features learned by the convolutional network. Results are given on a large generic object recognition task with six categories (human figures, four-legged animals, airplanes, trucks, cars, and "none of the above"), with multiple instances of each object category under various poses, illuminations, and backgrounds. On the test set, which contains different object instances than the training set, an SVM alone yields a 43.3% error rate, a convolutional net alone yields 7.2% and an SVM on top of features produced by the convolutional net yields 5.9%.

We describe a novel unsupervised method for learning sparse, overcomplete features. The model uses a linear encoder, and a linear decoder preceded by a sparsifying non-linearity that turns a code vector into a quasi-binary sparse code vector. Given an input, the optimal code minimizes the distance between the output of the decoder and the input patch while being as similar as possible to the encoder output. Learning proceeds in a two-phase EM-like fashion: (1) compute the minimum-energy code vector, (2) adjust the parameters of the encoder and decoder so as to decrease the energy. The model produces "stroke detectors" when trained on handwritten numerals, and Gabor-like filters when trained on natural image patches. Inference and learning are very fast, requiring no preprocessing, and no expensive sampling. Using the proposed unsupervised method to initialize the first layer of a convolutional network, we achieved an error rate slightly lower than the best reported result on the MNIST dataset. Finally, an extension of the method is described to learn topographical filter maps.

We present a method for training a similarity metric from data. The method can be used for recognition or verification applications where the number of categories is very large and not known during training, and where the number of training samples for a single category is very small. The idea is to learn a function that maps input patterns into a target space such that the L/sub 1/ norm in the target space approximates the "semantic" distance in the input space. The method is applied to a face verification task. The learning process minimizes a discriminative loss function that drives the similarity metric to be small for pairs of faces from the same person, and large for pairs from different persons. The mapping from raw to the target space is a convolutional network whose architecture is designed for robustness to geometric distortions. The system is tested on the Purdue/AR face database which has a very high degree of variability in the pose, lighting, expression, position, and artificial occlusions such as dark glasses and obscuring scarves.

We describe a vision-based obstacle avoidance system for off-road mobile robots. The system is trained from end to end to map raw in put images to steering angles. It is trained in supervised mode to predict the steering angles provided by a human driver during training r uns collected in a wide variety of terrains, weather conditions, lighting conditions, and obstacle types. The robot is a 50cm off-road truck, with two f orwardpointing wireless color cameras. A remote computer process es the video and controls the robot via radio. The learning system is a lar ge 6-layer convolutional network whose input is a single left/right pair of unprocessed low-resolution images. The robot exhibits an excellent ability to detect obstacles and navigate around them in real time at speeds of 2 m/s.

We describe a trainable system for analyzing videos of developing C. elegans embryos. The system automatically detects, segments, and locates cells and nuclei in microscopic images. The system was designed as the central component of a fully automated phenotyping system. The system contains three modules 1) a convolutional network trained to classify each pixel into five categories: cell wall, cytoplasm, nucleus membrane, nucleus, outside medium; 2) an energy-based model, which cleans up the output of the convolutional network by learning local consistency constraints that must be satisfied by label images; 3) a set of elastic models of the embryo at various stages of development that are matched to the label images.

We describe a novel method for real-time, simultaneous multi-view face detection and facial pose estimation. The method employs a convolu- tional network to map face images to points on a manifold, parametrized by pose, and non-face images to points far from that manifold. This network is trained by optimizing a loss function of three variables: im- age, pose, and face/non-face label. We test the resulting system, in a single configuration, on three standard data sets one for frontal pose, one for rotated faces, and one for profiles and find that its performance on each set is comparable to previous multi-view face detectors that can only handle one form of pose variation. We also show experimentally that the system's accuracy on both face detection and pose estimation is improved by training for the two tasks together. 1 Introduction The detection of human faces in natural images and videos is a key component in a wide variety of applications of human-computer interaction, search and indexing, security, and surveillance. Many real-world applications would profit from multi-view detectors that can detect faces under a wide range of poses: looking left or right (yaw axis), up or down (pitch axis), or tilting left or right (roll axis). In this paper we describe a novel method that not only detects faces independently of their poses, but simultaneously estimates those poses. The system is highly-reliable, runs at near real time (5 frames per second on standard hardware), and is robust against variations in yaw (90), roll (45), and pitch (60). The method is motivated by the idea that multi-view face detection and pose estimation are so closely related that they should not be performed separately. The tasks are related in the sense that they must be robust against the same sorts of variation: skin color, glasses, facial hair, lighting, scale, expressions, etc. We suspect that, when trained together, each task can serve as an inductive bias for the other, yielding better generalization or requiring fewer training examples [2]. To exploit the synergy between these two tasks, we train a convolutional network to map face images to points on a face manifold, and non-face images to points far away from that manifold. The manifold is parameterized by facial pose. Conceptually, we can view the pose parameter as a latent variable that can be inferred through an energy-minimization process [4]. To train the machine we derive a new type of discriminative loss function that is tailored to such detection tasks. Previous Work: Learning-based approaches to face detection abound, including real-time methods [16], and approaches based on convolutional networks [15, 3]. Most multi-view systems take a view-based approach, which involves building separate detectors for differ- ent views and either applying them in parallel [10, 14, 13, 7] or using a pose estimator to select a detector [5]. Another approach is to estimate and correct in-plane rotations before applying a single pose-specific detector [12]. Closer to our approach is that of [8], in which a number of Support Vector Regressors are trained to approximate smooth functions, each of which has a maximum for a face at a particular pose. Another machine is trained to con- vert the resulting values to estimates of poses, and a third is trained to convert the values into a face/non-face score. The resulting system is very slow. 2 Integrating face detection and pose estimation To exploit the posited synergy between face detection and pose estimation, we must design a system that integrates the solutions to the two problems. We hope to obtain better results on both tasks, so this should not be a mere cascaded system in which the answer to one problem is used to assist in solving the other. Both answers must be derived from one underlying analysis of the input, and both tasks must be trained together. Our approach is to build a trainable system that can map raw images X to points in a low-dimensional space. In that space, we pre-define a face manifold F (Z) that we para- meterize by the pose Z. We train the system to map face images with known poses to the corresponding points on the manifold. We also train it to map non-face images to points far away from the manifold. Proximity to the manifold then tells us whether or not an image is a face, and projection to the manifold yields an estimate of the pose. Parameterizing the Face Manifold: We will now describe the details of the parameter- izations of the face manifold. Let's start with the simplest case of one pose parameter Z = , representing, say, yaw. If we want to preserve the natural topology and geometry of the problem, the face manifold under yaw variations in the interval [-90, 90] should be a half circle (with constant curvature). We embed this half-circle in a three-dimensional space using three equally-spaced shifted cosines. Fi() = cos( - i); i = 1, 2, 3; = [- , ] (1) 2 2 When we run the network on an image X, it outputs a vector G(X) with three components that can be decoded analytically into corresponding pose angle: 3 G = arctan i=1 i(X ) cos(i) 3 (2) G i=1 i(X ) sin(i) The point on the manifold closest to G(X) is just F (). The same idea can be applied to any number of pose parameters. Let us consider the set of all faces with yaw in [-90, 90] and roll in [-45, 45]. In an abstract way, this set is isomorphic to a portion of the surface of a sphere. Consequently, we encode the pose with the product of the cosines of the two angles: Fij(, ) = cos( - i) cos( - j); i, j = 1, 2, 3; (3) For convenience we rescale the roll angles to the range of [-90, 90]. With these parame- terizations, the manifold has constant curvature, which ensures that the effect of errors will be the same regardless of pose. Given nine components of the network's output Gij(X), we compute the corresponding pose angles as follows: cc = G G ij ij (X ) cos(i) cos(j ); cs = ij ij (X ) cos(i) sin(j ) sc = G G ij ij (X ) sin(i) cos(j ); ss = ij ij (X ) sin(i) sin(j ) (4) = 0.5(atan2(cs + sc, cc - ss) + atan2(sc - cs, cc + ss)) = 0.5(atan2(cs + sc, cc - ss) - atan2(sc - cs, cc + ss)) Note that the dimension of the face manifold is much lower than that of the embedding space. This gives ample space to represent non-faces away from the manifold. 3 Learning Machine To build a learning machine for the proposed approach we refer to the Minimum Energy Machine framework described in [4]. Energy Minimization Framework: We can view our system as a scalar-value function EW (Y, Z, X), where X and Z are as defined above, Y is a binary label (Y = 1 for face, Y = 0 for a non-face), and W is a parameter vector subject to learning. EW (Y, Z, X) can be interpreted as an energy function that measures the degree of compatibility between X, Z, Y . If X is a face with pose Z, then we want: EW (1, Z, X) EW (0, Z , X) for any pose Z , and EW (1, Z , X) EW (1, Z, X) for any pose Z = Z. Operating the machine consists in clamping X to the observed value (the image), and finding the values of Z and Y that minimize EW (Y, Z, X): (Y , Z) = argminY {Y }, Z{Z}EW (Y, Z, X) (5) where {Y } = {0, 1} and {Z} = [-90, 90][-45, 45] for yaw and roll variables. Although this inference process can be viewed probabilistically as finding the most likely configu- ration of Y and Z according to a model that attributes high probabilities to low-energy configurations (e.g. a Gibbs distribution), we view it as a non probabilistic decision mak- ing process. In other words, we make no assumption as to the finiteness of integrals over {Y }and {Z}that would be necessary for a properly normalized probabilistic model. This gives us considerable flexibility in the choice of the internal architecture of EW (Y, Z, X). Our energy function for a face EW (1, Z, X) is defined as the distance between the point produced by the network GW (X) and the point with pose Z on the manifold F (Z): EW (1, Z, X) = GW (X) - F (Z) (6) The energy function for a non-face EW (0, Z, X) is equal to a constant T that we can interpret as a threshold (it is independent of Z and X). The complete energy function is: EW (Y, Z, X) = Y GW (X) - F (Z) + (1 - Y )T (7) The architecture of the machine is depicted in Figure 1. Operating this machine (find- ing the output label and pose with the smallest energy) comes down to first finding: Z = argminZ{Z}||GW (X) - F (Z)||, and then comparing this minimum distance, GW (X) - F (Z) , to the threshold T . If it smaller than T , then X is classified as a face, otherwise X is classified as a non-face. This decision is implemented in the architecture as a switch, that depends upon the binary variable Y . Convolutional Network: We employ a Convolutional Network as the basic architecture for the GW (X) image-to-face-space mapping function. Convolutional networks [6] are "end- to-end" trainable system that can operate on raw pixel images and learn low-level features and high-level representation in an integrated fashion. Convolutional nets are advantageous because they easily learn the types of shift-invariant local features that are relevant to image recognition; and more importantly, they can be replicated over large images (swept over every location) at a fraction of the cost of replicating more traditional classifiers [6]. This is a considerable advantage for building real-time systems. We employ a network architecture similar to LeNet5 [6]. The difference is in the number of maps. In our architecture we have 8 feature maps in the bottom convolutional and subsampling layers and 20 maps in the next two layers. The last layer has 9 outputs to encode two pose parameters. Training with a Discriminative Loss Function for Detection: We define the loss function as follows: 1 1 L(W ) = L L |S 1(W, Zi, X i) + 0(W, X i) (8) 1| |S iS 0| 1 iS0 Figure 1: Architecture of the Minimum Energy Machine. where S1is the set of training faces, S0the set of non-faces, L1(W, Zi, Xi) and L0(W, Xi) are loss functions for a face sample (with a known pose) and non-face, respectively1. The loss L(W ) should be designed so that its minimization for a particular positive training sample (Xi, Zi, 1), will make EW (1, Zi, Xi) < EW (Y, Z, Xi) for Y = Y i or Z = Zi. To satisfy this, it is sufficient to make EW (1, Zi, Xi) < EW (0, Z, Xi). For a particu- lar negative training sample (Xi, 0), minimizing the loss should make EW (1, Z, Xi) > EW (0, Z, Xi) = T for any Z. To satisfy this, it is sufficient to make EW (1, Z, Xi) > T . Let W be the current parameter value, and W be the parameter value after an update caused by a single sample. To cause the machine to achieve the desired behavior, we need the parameter update to decrease the difference between the energy of the desired label and the energy of the undesired label. In our case, since EW (0, Z, X) = T is constant, the following condition on the update is sufficient to ensure the desired behavior: Condition 1. for a face example (X, Z, 1), we must have: EW (1, Z, X) < EW (1, Z, X) For a non-face example (X, 1), we must have: EW (1, Z, X) > EW (1, Z, X) We choose the following forms for L1 and L0: L1(W, 1, Z, X) = EW (1, Z, X)2; L0(W, 0, X) = K exp[-E(1, Z, X)] (9) where K is a positive constant. Next we show that minimizing (9) with an incremental gradient-based algorithm will satisfy condition 1. With gradient-based optimization algorithms, the parameter update formula is of the form: W = W - W = -A L W . where A is a judiciously chosen symmetric positive semi-definite matrix, and is a small positive constant. For Y = 1 (face): An update step will change the parameter by W = -A EW (1,Z,X)2 = W -2EW (1, Z, X)A EW (1,Z,X) W . To first order (for small values of ), the resulting change in EW (1, Z, X) is given by: E T T W (1, Z, X ) E E W = -2E W (1, Z, X ) A W (1, Z, X ) < 0 W W (1, Z, X ) W W because EW (1, Z, X) > 0 (it's a distance), and the quadratic form is positive. Therefore EW (1, Z, X) < EW (1, Z, X). 1Although face samples whose pose is unknown can easily be accommodated, we will not discuss this possibility here. 100 100 95 95 90 90 85 85 80 80 75 75 70 70 65 65 Percentage of faces detected 60 Pose + detection 60 Pose + detection Detection only Pose only 55 Percentage of yaws correctly estimated 55 50 50 0 2 4 6 8 10 12 14 16 18 20 0 5 10 15 20 25 30 False positive rate Yaw-error tolerance (degrees) Figure 2: Synergy test. Left: ROC curves for the pose-plus-detection and detection-only networks. Right: frequency with which the pose-plus-detection and pose-only networks correctly estimated the yaws within various error tolerances. For Y = 0 (non-face): An update step will change the parameter by W = -A K exp[-E(1,Z,X)] = K exp[-E W W (1, Z, X )] EW (1,Z,X) W . To first order (for small values of ), the resulting change in EW (1, Z, X) is given by: E T T W (1, Z, X ) E E W = K exp[-E W (1, Z , X ) A W (1, Z , X ) > 0 W W (1, Z , X )] W W Therefore EW (1, Z, X) > EW (1, Z, X). Running the Machine: Our detection system works on grayscale images and it applies the network to each image at a range of scales, stepping by a factor of 2. The network is replicated over the image at each scale, stepping by 4 pixels in x and y (this step size is a consequence of having two, 2x2 subsampling layers). At each scale and location, the network outputs are compared to the closest point on the manifold, and the system collects a list of all instances closer than our detection threshold. Finally, after examining all scales, the system identifies groups of overlapping detections in the list and discards all but the strongest (closest to the manifold) from each group. No attempt is made to combine detections or apply any voting scheme. We have implemented the system in C. The system can detect, locate, and estimate the pose of faces that are between 40 and 250 pixels high in a 640 480 image at roughly 5 frames per second on a 2.4GHz Pentium 4. 4 Experiments and results Using the above architecture, we built a detector to locate faces and estimate two pose parameters: yaw from left to right profile, and in-plane rotation from -45 to 45 degrees. The machine was trained to be robust against pitch variation. In this section, we first describe the training regimen for this network, and then give the results of two sets of experiments. The first set of experiments tests whether training for the two tasks together improves performance on both. The second set allows comparisons between our system and other published multi-view detectors. Training: Our training set consisted of 52, 850, 32x32-pixel faces from natural images collected at NEC Labs and hand annotated with appropriate facial poses (see [9] for a description of how the annotation was done). These faces were selected from a much larger annotated set to yield a roughly uniform distribution of poses from left profile to right profile, with as much variation in pitch as we could obtain. Our initial negative training data consisted of 52, 850 image patches chosen randomly from non-face areas of a variety of images. For our second set of tests, we replaced half of these with image patches obtained by running the initial version of the detector on our training images and collecting false detections. Each training image was used 5 times during training, with random variations 100 100 95 95 90 90 85 85 80 80 75 75 70 70 65 65 Frontal Percentage of faces detected 60 Rotated in plane 60 In-plane rotation Profile Yaw 55 Percentage of poses correctly estimated 55 50 50 0 0.5 1 1.5 2 2.5 3 3.5 4 4.5 5 0 5 10 15 20 25 30 False positives per image Pose-error tolerance (degrees) Figure 3: Results on standard data sets. Left: ROC curves for our detector on the three data sets. The x axis is the average number of false positives per image over all three sets, so each point corresponds to a single detection threshold. Right: frequency with which yaw and roll are estimated within various error tolerances. in scale (from x 2 to x(1 + 2)), in-plane rotation (45), brightness (20), contrast (from 0.8 to 1.3). To train the network, we made 9 passes through this data, though it mostly converged after about the first 6 passes. Training was performed using LUSH [1], and the total training time was about 26 hours on a 2Ghz Pentium 4. At the end of training, the network had converged to an equal error rate of 5% on the training data and 6% on a separate test set of 90,000 images. Synergy tests: The goal of the synergy test was to verify that both face detection and pose estimation benefit from learning and running in parallel. To test this claim we built three networks with almost identical architectures, but trained to perform different tasks. The first one was trained for simultaneous face detection and pose estimation (combined), the second was trained for detection only and the third for pose estimation only. The "detection only" network had only one output for indicating whether or not its input was a face. The "pose only" network was identical to the combined network, but trained on faces only (no negative examples). Figure 2 shows the results of running these networks on our 10,000 test images. In both these graphs, we see that the pose-plus-detection network had better performance, confirming that training for each task benefits the other. Standard data sets: There is no standard data set that tests all the poses our system is designed to detect. There are, however, data sets that have been used to test more restricted face detectors, each set focusing on a particular variation in pose. By testing a single detector with all of these sets, we can compare our performance against published systems. As far as we know, we are the first to publish results for a single detector on all these data sets. The details of these sets are described below: MIT+CMU [14, 11] 130 images for testing frontal face detectors. We count 517 faces in this set, but the standard tests only use a subset of 507 faces, because 10 faces are in the wrong pose or otherwise not suitable for the test. (Note: about 2% of the faces in the standard subset are badly-drawn cartoons, which we do not intend our system to detect. Nevertheless, we include them in the results we report.) TILTED [12] 50 images of frontal faces with in-plane rotations. 223 faces out of 225 are in the standard subset. (Note: about 20% of the faces in the standard subset are outside of the 45 rotation range for which our system is designed. Again, we still include these in our results.) PROFILE [13] 208 images of faces in profile. There seems to be some disagreement about the number of faces in the standard set of annotations: [13] reports using 347 faces of the 462 that we found, [5] reports using 355, and we found 353 annotations. However, these discrepencies should not significantly effect the reported results. We counted a face as being detected if 1) at least one detection lay within a circle centered on the midpoint between the eyes, with a radius equal to 1.25 times the distance from that point to the midpoint of the mouth, and 2) that detection came at a scale within a factor of Figure 4: Some example face detections. Each white box shows the location of a detected face. The angle of each box indicates the estimated in-plane rotation. The black crosshairs within each box indicate the estimated yaw. Data set TILTED PROFILE MIT+CMU False positives per image 4.42 26.90 .47 3.36 .50 1.28 Our detector 90% 97% 67% 83% 83% 88% Jones & Viola [5] (tilted) 90% 95% x x Jones & Viola [5] (profile) x 70% 83% x Rowley et al [11] 89% 96% x x Schneiderman & Kanade [13] x 86% 93% x Table 1: Comparisons of our results with other multi-view detectors. Each column shows the detec- tion rates for a given average number of false positives per image (these rates correspond to those for which other authors have reported results). Results for real-time detectors are shown in bold. Note that ours is the only single detector that can be tested on all data sets simultaneously. two of the correct scale for the face's size. We counted a detection as a false positive if it did not lie within this range for any of the faces in the image, including those faces not in the standard subset. The left graph in Figure 3 shows ROC curves for our detector on the three data sets. Figure 4 shows a few results on various poses. Table 1 shows our detection rates compared against other systems for which results were given on these data sets. The table shows that our results on the TILTED and PROFILE sets are similar to those of the two Jones & Viola detectors, and even approach those of the Rowley et al and Schneiderman & Kanade non- real-time detectors. Those detectors, however, are not designed to handle all variations in pose, and do not yield pose estimates. The right side of Figure 3 shows our performance at pose estimation. To make this graph, we fixed the detection threshold at a value that resulted in about 0.5 false positives per image over all three data sets. We then compared the pose estimates for all detected faces (including those not in the standard subsets) against our manual pose annotations. Note that this test is more difficult than typical tests of pose estimation systems, where faces are first localized by hand. When we hand-localize these faces, 89% of yaws and 100% of in-plane rotations are correctly estimated to within 15.

We consider situations where training data is abundant and computing resources are comparatively scarce. We argue that suitably designed on- line learning algorithms asymptotically outperform any batch learning algorithm. Both theoretical and experimental evidences are presented.

We have implemented a real time front end for detecting voiced speech and estimating its fundamental frequency. The front end performs the signal processing for voice-driven agents that attend to the pitch contours of human speech and provide continuous audiovisual feedback. The al- gorithm we use for pitch tracking has several distinguishing features: it makes no use of FFTs or autocorrelation at the pitch period; it updates the pitch incrementally on a sample-by-sample basis; it avoids peak picking and does not require interpolation in time or frequency to obtain high res- olution estimates; and it works reliably over a four octave range, in real time, without the need for postprocessing to produce smooth contours. The algorithm is based on two simple ideas in neural computation: the introduction of a purposeful nonlinearity, and the error signal of a least squares fit. The pitch tracker is used in two real time multimedia applica- tions: a voice-to-MIDI player that synthesizes electronic music from vo- calized melodies, and an audiovisual Karaoke machine with multimodal feedback. Both applications run on a laptop and display the user’s pitch scrolling across the screen as he or she sings into the computer.

How can we turn the description of a digital (i.e. electronically produced) document into something that is efficient for multi-layer raster formats? It is first shown that a foreground/background segmentation without overlapping foreground components can be more efficient for viewing or printing. Then, a new algorithm that prevents overlaps between foreground components while optimizing both the document quality and compression ratio is derived from the minimum description length (MDL) criterion. This algorithm makes the DjVu compression format significantly, more efficient on electronically produced documents. Comparisons with other formats are provided.

DjVu is an image compression technique specifically geared towards the compression of scanned documents in color at high resolution. Typical color magazine pages scanned at 300 dpi are compressed to between 40 and 80 kBytes, or 5 to 10 times smaller than with JPEG for a similar level of subjective quality. The foreground layer, which contains the text and drawings and requires high spatial resolution, is separated from the background layer, which contains pictures and backgrounds and requires less resolution. The foreground is compressed with a bi-tonal image compression technique that takes advantage of character shape similarities. The background is compressed with a new progressive, wavelet-based compression method. A real-time, memory-efficient version of the decoder is available as a plug-in for popular Web browsers.

We present a new image compression technique called "DjVu" that is specifically geared towards the compression of scanned documents in color at high resolution. With DjVu, a magazine page in color at 300 dpi typically occupies between 40 KB and 80 KB, approximately 5 to 10 times better than JPEG for a similar level of readability. Using a combination of hidden Markov model techniques and MDL-driven heuristics, DjVu first classifies each pixel in the image as either foreground (text, drawings) or background (pictures, photos, paper texture). The pixel categories form a bitonal image which is compressed using a pattern matching technique that takes advantage of the similarities between character shapes. A progressive, wavelet-based compression technique, combined with a masking algorithm, is then used to compress the foreground and background images at lower resolutions while minimizing the number of bits spent on the pixels that are not visible in the foreground and background planes. Encoders, decoders, and real-time, memory efficient plug-ins for various web browsers are available for all the major platforms.

The challenge of multimedia processing is to provide services that seamlessly integrate text, sound, image, and video information and to do it in a way that preserves the ease of use and interactivity of conventional plain old telephone service (POTS) telephony. To achieve this goal, there are a number of technological problems that must be considered, including: compression and coding of multimedia signals, including algorithmic issues, standards issues, and transmission issues; synthesis and recognition of multimedia signals, including speech, images, handwriting, and text; organization, storage, and retrieval of multimedia signals, including the appropriate method and speed of delivery, resolution, and quality of service; access methods to the multimedia signal, including spoken natural language interfaces, agent interfaces, and media conversion tools; searching by text, speech, and image queries; browsing by accessing the text, by voice, or by indexed images. In each of these areas, a great deal of progress has been made in the past few years, driven in part by the relentless growth in multimedia personal computers and in part by the promise of broad-band access from the home and from wireless connections. Standards have also played a key role in driving new multimedia services, both on the POTS network and on the Internet. It is the purpose of this paper to review the status of the technology in each of the areas listed above and to illustrate current capabilities by describing several multimedia applications that have been implemented at AT&T Labs over the past several years.

Multilayer neural networks trained with the back-propagation algorithm constitute the best example of a successful gradient based learning technique. Given an appropriate network architecture, gradient-based learning algorithms can be used to synthesize a complex decision surface that can classify high-dimensional patterns, such as handwritten characters, with minimal preprocessing. This paper reviews various methods applied to handwritten character recognition and compares them on a standard handwritten digit recognition task. Convolutional neural networks, which are specifically designed to deal with the variability of 2D shapes, are shown to outperform all other techniques. Real-life document recognition systems are composed of multiple modules including field extraction, segmentation recognition, and language modeling. A new learning paradigm, called graph transformer networks (GTN), allows such multimodule systems to be trained globally using gradient-based methods so as to minimize an overall performance measure. Two systems for online handwriting recognition are described. Experiments demonstrate the advantage of global training, and the flexibility of graph transformer networks. A graph transformer network for reading a bank cheque is also described. It uses convolutional neural network character recognizers combined with global training techniques to provide record accuracy on business and personal cheques. It is deployed commercially and reads several million cheques per day.

Discusses coding standards for still images and motion video. We first briefly discuss standards already in use, including: Group 3 and Group 4 for bilevel fax images; JPEG for still color images; and H.261, H.263, MPEG-1, and MPEG-2 for motion video. We then cover newly emerging standards such as JBIG1 and JBIG2 for bilevel fax images, JPEG-2000 for still color images, and H.263+ and MPEG-4 for motion video. Finally, we describe some directions beyond the standards such as hybrid coding of graphics/photo images, MPEG-7 for multimedia metadata, and possible new technologies.

We propose a new machine learning paradigm called Graph Transformer Networks that extends the applicability of gradient-based learning algorithms to systems composed of modules that take graphs as inputs and produce graphs as output. Training is performed by computing gradients of a global objective function with respect to all the parameters in the system using a kind of back-propagation procedure. A complete check reading system based on these concepts is described. The system uses convolutional neural network character recognizers, combined with global training techniques to provide record accuracy on business and personal checks. It is presently deployed commercially and reads million of checks per month.

We propose a new machine learning paradigm called multilayer graph transformer network that extends the applicability of gradient-based learning algorithms to systems composed of modules that take graphs as input and produce graphs as output. A complete check reading system based on this concept is described. The system combines convolutional neural network character recognizers with graph-based stochastic models trained cooperatively at the document level. It is deployed commercially and reads million of business and personal checks per month with record accuracy.

The challenge of multimedia processing is to seamlessly integrate text, sound, image, and video information into a single communications channel, and to do it in a way that provides high quality communications while preserving the ease-of-use and interactivity of conventional telephony. There are a number of technology drivers that are pushing the technology forward, as well as a number of technological problems that must be overcome before multimedia becomes as ubiquitous as voiceband telephony. A key issue with any practical multimedia system has to do with standards that insure connectivity between customers and a range of service providers. Multimedia processing is an area of communications that is rapidly evolving. However, a number of interesting and important multimedia communications applications have evolved over the past several years, and some of these applications are described.

A system for discriminative feature and model design is presented for automatic speech recognition. Training based on minimum classification error with a single objective function is applied for designing a set of parallel networks performing feature transformation and a set of hidden Markov models performing speech recognition. This paper compares the use of linear and non-linear functional transformations when applied to conventional recognition features, such as spectrum or cepstrum. It also provides a framework for integrated feature and model training when using class-specific transformations. Experimental results on telephone-based connected digit recognition are presented. 1. INTRODUCTION Improving the performance of hidden Markov model (HMM) based automatic speech recognition (ASR) systems has been a central issue that has dominated the entire field of speech recognition during the past two decades. One effort to improving HMMs has been by extending the training paradigm beyo...

We introduce a new approach for on-line recognition of handwritten words written in unconstrained mixed style. The preprocessor performs a word-level normalization by fitting a model of the word structure using the EM algorithm. Words are then coded into low resolution "annotated images" where each pixel contains information about trajectory direction and curvature. The recognizer is a convolution network that can be spatially replicated. From the network output, a hidden Markov model produces word scores. The entire system is globally trained to minimize word-level errors.

We introduce a new approach to normalizing words written with an electronic stylus that applies to all styles of handwriting (upper case, lower case, printed, cursive, or mixed). A geometrical model of the word spatial structure is fitted to the pen trajectory using the expectation-maximisation algorithm. The fitting process maximizes the likelihood of the trajectory given the model and a set a priors on its parameters. The method was evaluated and integrated to a recognition system that combines neural networks and hidden Markov models.

This paper compares the performance of several classifier algorithms on a standard database of handwritten digits. We consider not only raw accuracy, but also training time, recognition time, and memory requirements. When available, we report measurements of the fraction of patterns that must be rejected so that the remaining patterns have misclassification rates less than a given threshold.

We introduce a new approach for online recognition of handwritten words written in unconstrained mixed style. Words are represented by low resolution "annotated images" where each pixel contains information about trajectory direction and curvature. The recognizer is a convolutional network which can be spatially replicated. From the network output, a hidden Markov model produces word scores. The entire system is globally trained to minimize word-level errors.

Memory-based classification algorithms such as radial basis functions or K-nearest neighbors often rely on simple distances (Euclidean distance, Hamming distance, etc.), which are rarely meaningful on pattern vectors. More complex better suited distance measures are often expensive and rather ad-hoc. We propose a new distance measure which: 1) can be made locally invariant to any set of transformations of the input; and 2) can be computed efficiently. We tested the method on large handwritten character databases provided by the US Post Office and NIST. Using invariances with respect to translation, rotation, scaling, skewing and line thickness, the method outperformed all other systems on a small (less than 10,000 patterns) database and was competitive on our largest (60,000 patterns) database.

We compare the performance of three types of neural network-based ensemble techniques to that of a single neural network. The ensemble algorithms are two versions of boosting and committees of neural networks trained independently. For each of the four algorithms, we experimentally determine the test and training error curves in an optical character recognition (OCR) problem as both a function of training set size and computational cost using three architectures. We show that a single machine is best for small training set size while for large training set size some version of boosting is best. However, for a given computational cost, boosting is always best. Furthermore, we show a surprising result for the original boosting algorithm: namely, that as the training set size increases, the training error decreases until it asymptotes to the test error rate. This has potential implications in the search for better training algorithms.

A method for measuring the capacity of learning machines is described. The method is based on fitting a theoretically derived function to empirical measurements of the maximal difference between the error rates on two separate data sets of varying sizes. Experimental measurements of the capacity of various types of linear classifiers are presented.

This paper describes the development of an algorithm for verification of signatures written on a touch-sensitive pad. The signature verification algorithm is based on an artificial neural network. The novel network presented here, called a “Siamese” time delay neural network, consists of two identical networks joined at their output. During training the network learns to measure the similarity between pairs of signatures. When used for verification, only one half of the Siamese network is evaluated. The output of this half network is the feature vector for the input signature. Verification consists of comparing this feature vector with a stored feature vector for the signer. Signatures closer than a chosen threshold to this stored representation are accepted, all other signatures are rejected as forgeries. System performance is illustrated with experiments performed in the laboratory.

We describe a method, “Shortest Path Segmentation” (SPS), which combines dynamic programming and a neural net recognizer for segmenting and recognizing character strings. We describe the application of this method to two problems: recognition of handwritten ZIP Codes, and recognition of handwritten words. For the ZIP Codes, we also used the method to automatically segment the images during training: the dynamic programming stage both performs the segmentation and provides inputs and desired outputs to the neural network. Results are reported for a test set of 2642 unsegmented handwritten 212 dpi binary ZIP Code (5- and 9-digit) images. For handwritten word recognition, we combined SPS with a “Space Displacement Neural Network” approach, in which a single-character-recognition network is extended over the entire word image, and in which SPS techniques are then used to rank order a given lexicon. We report results on a test set of 3000 300 ppi gray scale word images, extracted from images of live mail pieces, for lexicons of size 10, 100, and 1000. Representing the problem as a graph as proposed in this paper has advantages beyond the efficient finding of the final optimal segmentation, or the automatic segmentation of images during training. We can also easily extend the technique to generate K “runner up” answers (for example, by finding the K shortest paths). This paper will also describe applications of some of these ideas.

In many machine learning applications, one has not only training data but also some high-level information about certain invariances that the system should exhibit. In character recognition, for example, the answer should be invariant with respect to small spatial distortions in the input images (translations, rotations, scale changes, etcetera). The authors have implemented a scheme that minimizes the derivative of the classifier outputs with respect to distortion operators. This not only produces tremendous speed advantages, but also provides a powerful language for specifying what generalizations the network can perform.>

In order to generalize from a training set to a test set, it is desirable that small changes in the input space of a pattern do not change the output components. This can be done by forcing this behavior as part of the training algorithm. This is done in double backpropagation by forming an energy function that is the sum of the normal energy term found in backpropagation and an additional term that is a function of the Jacobian. Significant improvement is shown with different architectures and different test sets, especially with architectures that had previously been shown to have very good performance when trained using backpropagation. It is shown that double backpropagation, as compared to backpropagation, creates weights that are smaller, thereby causing the output of the neurons to spend more time in the linear region.

A neural network with 136000 connections for recognition of handwritten digits has been implemented using a mixed analog/digital neural network chip. The neural network chip is capable of processing 1000 characters/s. The recognition system has essentially the same rate (5%) as a simulation of the network with 32-b floating-point precision.

An application of back-propagation networks to handwritten zip code recognition is presented. Minimal preprocessing of the data is required, but the architecture of the network is highly constrained and specifically designed for the task. The input of the network consists of size-normalized images of isolated digits. The performance on zip code digits provided by the US Postal Service is 92% recognition, 1% substitution, and 7% rejects. Structured neural networks can be viewed as statistical methods with structure which bridge the gap between purely statistical and purely structural methods.>

Hardware architectures for character recognition are discussed, and choices for possible circuits are outlined. An advanced (and working) reconfigurable neural-net chip that mixes analog and digital processing is described. It is found that different approaches to image recognition often lead to neural-net architectures that have limited connectivity and repeated use of the same set of weights. This architecture is ideal for time-multiplexing (a combined parallel-series processing) on hardware systems that would be too small to evaluate the entire network in parallel. To make this process efficient, a chip needs to have shift registers to format the input data and additional registers to store intermediate results. Within this framework, it is possible to design chips that have broad utility, large connection capacity, and high speed. This was demonstrated by a new chip with 32000 reconfigurable connections

The ability of learning networks to generalize can be greatly enhanced by providing constraints from the task domain. This paper demonstrates how such constraints can be integrated into a backpropagation network through the architecture of the network. This approach has been successfully applied to the recognition of handwritten zip code digits provided by the U.S. Postal Service. A single network learns the entire recognition operation, going from the normalized image of the character to the final classification.