This paper revisits the canonical concept of learning structured representations without label supervision by eigendecomposition. Yet, unlike prior spectral methods such as Laplacian Eigenmap which operate in a nonparametric manner, we aim to parametrically model the principal eigenfunctions of an integral operator defined by a kernel and a data distribution using a neural network for enhanced scalability and reasonable out-of-sample generalization. To achieve this goal, we first present a new series of objective functions that generalize the EigenGame Gemp et al. 2020 to function space for learning neural eigenfunctions. We then show that, when the similarity metric is derived from positive relations in a data augmentation setup, a representation learning objective function that resembles those of popular self-supervised learning methods emerges, with an additional symmetry-breaking property for producing structured representations where features are ordered by importance. We call such a structured, adaptive-length deep representation Neural Eigenmap. We demonstrate using Neural Eigenmap as adaptive-length codes in image retrieval systems. By truncation according to feature importance, our method requires up to $16\times$16× shorter representation length than leading self-supervised learning ones to achieve similar retrieval performance. We further apply our method to graph data and report strong results on a node representation learning benchmark with more than one million nodes.

Specializing large language models (LLMs) for local deployment in domain-specific use cases is necessary for strong performance while meeting latency and privacy constraints.However, conventional task-specific adaptation approaches do not show simultaneous memory saving and inference speedup at deployment time.Practical compression techniques like quantization and pruning require dedicated hardware or kernel support to achieve measured inference speedup.We develop TRIM-LLM based on the layer-wise specialization phenomenon we empirically observed and verified on contemporary LLMs.TRIMLLM reduces the depth of LLMs via progressive layer dropping.We show it retains LLMs' capacity in specific domains and achieves inference speedup irrespective of hardware and deep learning frameworks.We evaluated TRIM-LLM on LLMs of various sizes for inference; models adapted on medical, legal, and financial datasets all demonstrate 2.1 -5.7× inference speedup on consumer GPUs and up to 3.1× speedup on A100 when compared to state-of-the-art model compression algorithms, with no loss in accuracy at 50∼60% model compression ratio.Our code is available at https://github.com/snyhlxde1/TrimLLM.

Training large language models for long context understanding faces the challenge of data shortage. Previous data engineering approaches mechanically concatenate short documents, which may create many pseudo long documents but raise concerns about data quality. In this paper, we study the core attribute of high quality data for long context training, and provide a data pipeline, LongPack, to scale such data. We found that long distance referrals, which occur in natural long documents, are crucial for long-context training. However, simply concatenating short documents does not reliably generate these relations. We further show that the density of long-distance referrals, which is higher in longer documents, has a key role in training efficiency, making previous upsampling methods suboptimal. To enrich long documents, we propose LongPack, a data pipeline that constructs long documents by packing shorter ones based on referral relationships. Specifically, for web pages, which are the primary source for language model training, we found hyper-link a native signal for such a relation. By packing web pages through their hyper-link connection, we can create longer, high-quality documents. Our experiments demonstrate that LongPackis highly scalable, generating a corpus of long documents equivalent in size to an entire pretraining dataset using just 0.5% root documents. Furthermore, the constructed documents have a ‘near-natural’ quality as innate long documents for long context training, reaching a 32.7% higher score than previous state-of-the-art methods.

Evaluating the reasoning abilities of large language models (LLMs) is challenging. Existing benchmarks often depend on static datasets, which are vulnerable to data contamination and may get saturated over time, or on binary live human feedback that conflates reasoning with other abilities. As the most prominent dynamic benchmark, Chatbot Arena evaluates open-ended questions in real-world settings, but lacks the granularity in assessing specific reasoning capabilities. We introduce GameArena, a dynamic benchmark designed to evaluate LLM reasoning capabilities through interactive gameplay with humans. GameArena consists of three games designed to test specific reasoning capabilities (e.g., deductive and inductive reasoning), while keeping participants entertained and engaged. We analyze the gaming data retrospectively to uncover the underlying reasoning processes of LLMs and measure their fine-grained reasoning capabilities. We collect over 2000 game sessions and provide detailed assessments of various reasoning capabilities for five state-of-the-art LLMs. Our user study with 100 participants suggests that GameArena improves user engagement compared to Chatbot Arena. For the first time, GameArena enables the collection of step-by-step LLM reasoning data in the wild.

Diffusion Transformers (DiTs) with 3D full attention power state-of-the-art video generation, but suffer from prohibitive compute cost -- when generating just a 5-second 720P video, attention alone takes 800 out of 950 seconds of total inference time. This paper introduces sliding tile attention (STA) to address this challenge. STA leverages the observation that attention scores in pretrained video diffusion models predominantly concentrate within localized 3D windows. By sliding and attending over local spatial-temporal region, STA eliminates redundancy from full attention. Unlike traditional token-wise sliding window attention (SWA), STA operates tile-by-tile with a novel hardware-aware sliding window design, preserving expressiveness while being \emph{hardware-efficient}. With careful kernel-level optimizations, STA offers the first efficient 2D/3D sliding-window-like attention implementation, achieving 58.79\% MFU -- 7.17× faster than prior art methods. On the leading video DiT model, Hunyuan, it accelerates attention by 1.6–10x over FlashAttention-3, yielding a 1.36–3.53× end-to-end speedup with no or minimum quality loss.

Test-time reasoning algorithms such as chain-of-thought, self-consistency, and MCTS enhance LLM problem-solving but can wastefully generate many tokens without improving accuracy. At the same time, we observe that these algorithms exhibit answer stabilization: their intermediate solutions often cease to change after a certain point, and further investment of compute does not change their final answer. To quantify this phenomenon, we introduce Certaindex, an algorithm-agnostic metric measuring this evolving stability, signaling when further computation is unlikely to alter the final result. Certaindex is lightweight, can accelerate reasoning program inference via early exit, and further enables dynamic token allocation, gang scheduling, and many opportunities when integrated with real-world LLM serving systems. To quantify real-world benefits, we built Certaindex as a scheduler into Dynasor, our reasoning-aware LLM serving system, and demonstrate up to 50\% compute savings and 3.3$\times$ higher throughput in real workloads with no accuracy drop. Our code is available at https://github.com/hao-ai-lab/Dynasor.git

Reasoning models excel by generating long chain-of-thoughts, but decoding the resulting thousands of tokens is slow. Token-level specualtive decoding (SD) helps, but its benefit is capped, because the chance that an entire $\gamma$-token guess is correct falls exponentially as $\gamma$ grows. This means allocating more compute for longer token drafts faces an algorithmic ceiling -- making the speedup modest and hardware-agnostic. We raise this ceiling with lookahead reasoning, which exploits a second, step-level layer of parallelism. Our key insight is that reasoning models generate step-by-step, and each step needs only to be semantically correct, not exact token matching. In lookahead reasoning, a lightweight draft model proposes several future steps; the target model expands each proposal in one batched pass, and a verifier keeps semantically correct steps while letting the target regenerate any that fail. Token-level SD still operates within each reasoning step, so the two layers of parallelism multiply. We show lookahead reasoning lifts the peak speedup of SD both theoretically and empirically. Across GSM8K, AIME, and other benchmarks, lookahead reasoning improves the speedup of SD from 1.4x to 2.1x while preserving answer quality, and its speedup scales better with additional GPU throughput. Our code is available at https://github.com/hao-ai-lab/LookaheadReasoning

Scaling video diffusion transformers (DiTs) is limited by their quadratic 3D attention, even though most of the attention mass concentrates on a small subset of positions. We turn this observation into VSA, a trainable, hardware-efficient sparse attention that replaces full attention at both training and inference. In VSA, a lightweight coarse stage pools tokens into tiles and identifies high-weight critical tokens; a fine stage computes token-level attention only inside those tiles subjecting to block computing layout to ensure hard efficiency. This leads to a single differentiable kernel that trains end-to-end, requires no post-hoc profiling, and sustains 85\% of FlashAttention3 MFU. We perform a large sweep of ablation studies and scaling-law experiments by pretraining DiTs from 60M to 1.4B parameters. VSA reaches a Pareto point that cuts training FLOPS by 2.53$\times$ with no drop in diffusion loss. Retrofitting the open-source Wan2.1-1.3B model speeds up attention time by 6$\times$ and lowers end-to-end generation time from 31s to 18s with comparable quality, while for the 14B model, end-to-end generation time is reduced from 1274s to 576s. Furthermore, we introduce a preliminary study of Sparse-Distill, the first method to enable sparse attention and distillation concurrently, achieving 50.9x speed up for Wan-1.3B while maintaining quality. These results establish trainable sparse attention as a practical alternative to full attention and a key enabler for further scaling of video diffusion models. Code is available at https://github.com/hao-ai-lab/FastVideo.

Studying how people interact with large language models (LLMs) in real-world scenarios is increasingly important due to their widespread use in various applications. In this paper, we introduce LMSYS-Chat-1M, a large-scale dataset containing one million real-world conversations with 25 state-of-the-art LLMs. This dataset is collected from 210K unique IP addresses in the wild on our Vicuna demo and Chatbot Arena website. We offer an overview of the dataset's content, including its curation process, basic statistics, and topic distribution, highlighting its diversity, originality, and scale. We demonstrate its versatility through four use cases: developing content moderation models that perform similarly to GPT-4, building a safety benchmark, training instruction-following models that perform similarly to Vicuna, and creating challenging benchmark questions. We believe that this dataset will serve as a valuable resource for understanding and advancing LLM capabilities. The dataset is publicly available at https://huggingface.co/datasets/lmsys/lmsys-chat-1m.

Large Language Models (LLMs) have unlocked new capabilities and applications; however, evaluating the alignment with human preferences still poses significant challenges. To address this issue, we introduce Chatbot Arena, an open platform for evaluating LLMs based on human preferences. Our methodology employs a pairwise comparison approach and leverages input from a diverse user base through crowdsourcing. The platform has been operational for several months, amassing over 240K votes. This paper describes the platform, analyzes the data we have collected so far, and explains the tried-and-true statistical methods we are using for efficient and accurate evaluation and ranking of models. We confirm that the crowdsourced questions are sufficiently diverse and discriminating and that the crowd-sourced human votes are in good agreement with those of expert raters. These analyses collectively establish a robust foundation for the credibility of Chatbot Arena. Because of its unique value and openness, Chatbot Arena has emerged as one of the most referenced LLM leaderboards, widely cited by leading LLM developers and companies. The platform is publicly available at https://chat.lmsys.org.

Large language models (LLMs) have demonstrated remarkable performance, and organizations are racing to serve LLMs of varying sizes as endpoints for use-cases like chat, programming and search. However, efficiently serving multiple LLMs poses significant challenges for existing approaches due to varying popularity of LLMs. In the paper, we present MuxServe, a flexible spatial-temporal multiplexing system for efficient multiple LLM serving. The key insight behind is to colocate LLMs considering their popularity to multiplex memory resources, and leverage the characteristics of prefill and decoding phases to separate and flexibly colocate them to multiplex computation resources. MuxServe formally formulates the multiplexing problem, and proposes a novel placement algorithm and adaptive batch scheduling strategy to identify optimal colocations and maximize utilization. MuxServe designs a unified resource manager to enable flexible and efficient multiplexing. Evaluation results show that MuxServe can achieves up to $1.8\times$ higher throughput or processes $2.9\times$ more requests within $99\%$ SLO attainment. The code is available at: https://github.com/hao-ai-lab/MuxServe.

Autoregressive decoding of large language models (LLMs) is memory bandwidth bounded, resulting in high latency and significant wastes of the parallel processing power of modern accelerators. Existing methods for accelerating LLM decoding often require a draft model (e.g., speculative decoding), which is nontrivial to obtain and unable to generalize. In this paper, we introduce Lookahead decoding, an exact, parallel decoding algorithm that accelerates LLM decoding without needing auxiliary models or data stores. It allows trading per-step log(FLOPs) to reduce the number of total decoding steps, is more parallelizable on single or multiple modern accelerators, and is compatible with concurrent memory-efficient attention (e.g., FlashAttention). Our implementation of Lookahead decoding can speed up autoregressive decoding by up to 1.8x on MT-bench and 4x with strong scaling on multiple GPUs in code completion tasks. Our code is avialable at https://github.com/hao-ai-lab/LookaheadDecoding

Jacobi decoding shows promise for more efficient LLM inference as it breaks the sequential nature of the LLM decoding process and transforms it into more parallelizable computation. However, in practice, it achieves little speedup compared to traditional autoregressive (AR) decoding, primarily because Jacobi decoding seldom accurately predicts more than one token in a single fixed-point iteration step. To address this, we develop a new approach aimed at realizing fast convergence from any state to the fixed point in a Jacobi trajectory. This is accomplished by refining the target LLM to consistently predict the fixed point given any state as input. Extensive experiments demonstrate the effectiveness of our method, showing 2.4$\times$ to 3.4$\times$ improvements in generation speed while preserving generation quality across both domain-specific and open-domain benchmarks.

Speculative decoding is a pivotal technique to accelerate the inference of large language models (LLMs) by employing a smaller draft model to predict the target model’s outputs. However, its efficacy can be limited due to the low predictive accuracy of the draft model, particularly when faced with diverse text inputs and a significant capability gap between the draft and target models. We introduce online speculative decoding to address this challenge. The main idea is to continuously update the (multiple) draft model(s) on observed user query data. Adapting to query distribution mitigates the shifts between the training distribution of the draft model and the query distribution, enabling the draft model to more accurately predict the target model’s outputs. We develop a prototype of online speculative decoding based on knowledge distillation and evaluate it using both synthetic and real query data. The results show a substantial increase in the token acceptance rate by 0.1 to 0.65, bringing 1.42x to 2.17x latency reduction. Our code is available at https://github.com/LiuXiaoxuanPKU/OSD.

In Large Language Model (LLM) inference, the output length of an LLM request is typically regarded as not known a priori. Consequently, most LLM serving systems employ a simple First-come-first-serve (FCFS) scheduling strategy, leading to Head-Of-Line (HOL) blocking and reduced throughput and service quality. In this paper, we reexamine this assumption -- we show that, although predicting the exact generation length of each request is infeasible, it is possible to predict the relative ranks of output lengths in a batch of requests, using learning to rank. The ranking information offers valuable guidance for scheduling requests. Building on this insight, we develop a novel scheduler for LLM inference and serving that can approximate the shortest-job-first (SJF) schedule better than existing approaches. We integrate this scheduler with the state-of-the-art LLM serving system and show significant performance improvement in several important applications: 2.8x lower latency in chatbot serving and 6.5x higher throughput in synthetic data generation. Our code is available at https://github.com/hao-ai-lab/vllm-ltr.git

The quadratic complexity and weak length extrapolation of Transformers limits their ability to scale to long sequences, and while sub-quadratic solutions like linear attention and state space models exist, they empirically underperform Transformers in pretraining efficiency and downstream task accuracy. We introduce MEGALODON, an neural architecture for efficient sequence modeling with unlimited context length. MEGALODON inherits the architecture of MEGA (exponential moving average with gated attention), and further introduces multiple technical components to improve its capability and stability, including complex exponential moving average (CEMA), timestep normalization layer, normalized attention mechanism and pre-norm with two-hop residual configuration. In a controlled head-to-head comparison with LLAMA2, MEGALODON achieves better efficiency than Transformer in the scale of 7 billion parameters and 2 trillion training tokens. MEGALODON reaches a training loss of 1.70, landing mid-way between LLAMA2-7B (1.75) and LLAMA2-13B (1.67). This result is robust throughout a wide range of benchmarks, where MEGALODON consistently outperforms Transformers across different tasks, domains, and modalities.

Evaluating large language model (LLM) based chat assistants is challenging due to their broad capabilities and the inadequacy of existing benchmarks in measuring human preferences.To address this, we explore using strong LLMs as judges to evaluate these models on more open-ended questions.We examine the usage and limitations of LLM-as-a-judge, including position, verbosity, and self-enhancement biases, as well as limited reasoning ability, and propose solutions to mitigate some of them.We then verify the agreement between LLM judges and human preferences by introducing two benchmarks: MT-bench, a multi-turn question set; and Chatbot Arena, a crowdsourced battle platform.Our results reveal that strong LLM judges like GPT-4 can match both controlled and crowdsourced human preferences well, achieving over 80\% agreement, the same level of agreement between humans.Hence, LLM-as-a-judge is a scalable and explainable way to approximate human preferences, which are otherwise very expensive to obtain.Additionally, we show our benchmark and traditional benchmarks complement each other by evaluating several variants of LLaMA and Vicuna.The MT-bench questions, 3K expert votes, and 30K conversations with human preferences are publicly available at https://github.com/lm-sys/FastChat/tree/main/fastchat/llm_judge.

High throughput serving of large language models (LLMs) requires batching sufficiently many requests at a time. However, existing systems struggle because the key-value cache (KV cache) memory for each request is huge and grows and shrinks dynamically. When managed inefficiently, this memory can be significantly wasted by fragmentation and redundant duplication, limiting the batch size. To address this problem, we propose PagedAttention, an attention algorithm inspired by the classical virtual memory and paging techniques in operating systems. On top of it, we build vLLM, an LLM serving system that achieves (1) near-zero waste in KV cache memory and (2) flexible sharing of KV cache within and across requests to further reduce memory usage. Our evaluations show that vLLM improves the throughput of popular LLMs by 2--4× with the same level of latency compared to the state-of-the-art systems, such as FasterTransformer and Orca. The improvement is more pronounced with longer sequences, larger models, and more complex decoding algorithms. vLLM's source code is publicly available at https://github.com/vllm-project/vllm.

Scaling up model sizes can lead to fundamentally new capabilities in many machine learning (ML) tasks. However, training big models requires strong distributed system expertise to carefully design model-parallel execution strategies that suit the model architectures and cluster setups. In this paper, we develop AMP, a framework that automatically derives such strategies. AMP identifies a valid space of model parallelism strategies and efficiently searches the space for high-performed strategies, by leveraging a cost model designed to capture the heterogeneity of the model and cluster specifications. Unlike existing methods, AMP is specifically tailored to support complex models composed of uneven layers and cluster setups with more heterogeneous accelerators and bandwidth. We evaluate AMP on popular modelsand cluster setups from public clouds and show that AMP returns parallel strategies that match the expert-tuned strategies on typical cluster setups. On heterogeneous clusters or models with heterogeneous architectures, AMP finds strategies with 1.54$\times$ and 1.77$\times$ higher throughput than state-of-the-art model-parallel systems, respectively.

Clustering algorithms based on deep neural networks have been widely studied for image analysis. Most existing methods require partial knowledge of the true labels, namely, the number of clusters, which is usually not available in practice. In this article, we propose a Bayesian nonparametric framework, deep nonparametric Bayes (DNB), for jointly learning image clusters and deep representations in a doubly unsupervised manner. In doubly unsupervised learning, we are dealing with the problem of "unknown unknowns," where we estimate not only the unknown image labels but also the unknown number of labels as well. The proposed algorithm alternates between generating a potentially unbounded number of clusters in the forward pass and learning the deep networks in the backward pass. With the help of the Dirichlet process mixtures, the proposed method is able to partition the latent representations space without specifying the number of clusters a priori. An important feature of this work is that all the estimation is realized with an end-to-end solution, which is very different from the methods that rely on post hoc analysis to select the number of clusters. Another key idea in this article is to provide a principled solution to the problem of "trivial solution" for deep clustering, which has not been much studied in the current literature. With extensive experiments on benchmark datasets, we show that our doubly unsupervised method achieves good clustering performance and outperforms many other unsupervised image clustering methods.

Panoptic segmentation that unifies instance segmentation and semantic segmentation has recently attracted increasing attention. While most existing methods focus on designing novel architectures, we steer toward a different perspective: performing automated multi-loss adaptation (named Ada-Segment) on the fly to flexibly adjust multiple training losses over the course of training using a controller trained to capture the learning dynamics. This offers a few advantages: it bypasses manual tuning of the sensitive loss combination, a decisive factor for panoptic segmentation; allows to explicitly model the learning dynamics, and reconcile the learning of multiple objectives (up to ten in our experiments); with an end-to-end architecture, it generalizes to different datasets without the need of re-tuning hyperparameters or re-adjusting the training process laboriously. Our Ada-Segment brings 2.7% panoptic quality (PQ) improvement on COCO val split from the vanilla baseline, achieving the state-of-the-art 48.5% PQ on COCO test-dev split and 32.9% PQ on ADE20K dataset. The extensive ablation studies reveal the ever-changing dynamics throughout the training process, necessitating the incorporation of an automated and adaptive learning strategy as presented in this paper.

Model parallelism has become a necessity for training modern large-scale deep language models. In this work, we identify a new and orthogonal dimension from existing model parallel approaches: it is possible to perform pipeline parallelism within a single training sequence for Transformer-based language models thanks to its autoregressive property. This enables a more fine-grained pipeline compared with previous work. With this key idea, we design TeraPipe, a high-performance token-level pipeline parallel algorithm for synchronous model-parallel training of Transformer-based language models. We develop a novel dynamic programming-based algorithm to calculate the optimal pipelining execution scheme given a specific model and cluster configuration. We show that TeraPipe can speed up the training by 5.0x for the largest GPT-3 model with 175 billion parameters on an AWS cluster with 48 p3.16xlarge instances compared with state-of-the-art model-parallel methods. The code for reproduction can be found at https://github.com/zhuohan123/terapipe

In recent years, the pace of innovations in the fields of machine learning (ML) has accelerated, researchers in SysML have created algorithms and systems that parallelize ML training over multiple devices or computational nodes. As ML models become more structurally complex, many systems have struggled to provide all-round performance on a variety of models. Particularly, ML scale-up is usually underestimated in terms of the amount of knowledge and time required to map from an appropriate distribution strategy to the model. Applying parallel training systems to complex models adds nontrivial development overheads in addition to model prototyping, and often results in lower-than-expected performance. This tutorial identifies research and practical pain points in parallel ML training, and discusses latest development of algorithms and systems on addressing these challenges in both usability and performance. In particular, this tutorial presents a new perspective of unifying seemingly different distributed ML training strategies. Based on it, introduces new techniques and system architectures to simplify and automate ML parallelization. This tutorial is built upon the authors' years' of research and industry experience, comprehensive literature survey, and several latest tutorials and papers published by the authors and peer researchers.

Synchronization is a key step in data-parallel distributed machine learning (ML). Different synchronization systems and strategies perform differently, and to achieve optimal parallel training throughput requires synchronization strategies that adapt to model structures and cluster configurations. Existing synchronization systems often only consider a single or a few synchronization aspects, and the burden of deciding the right synchronization strategy is then placed on the ML practitioners, who may lack the required expertise. In this paper, we develop a model- and resource-dependent representation for synchronization, which unifies multiple synchronization aspects ranging from architecture, message partitioning, placement scheme, to communication topology. Based on this representation, we build an end-to-end pipeline, AutoSync, to automatically optimize synchronization strategies given model structures and resource specifications, lowering the bar for data-parallel distributed ML. By learning from low-shot data collected in only 200 trial runs, AutoSync can discover synchronization strategies up to 1.6x better than manually optimized ones. We develop transfer-learning mechanisms to further reduce the auto-optimization cost -- the simulators can transfer among similar model architectures, among similar cluster configurations, or both. We also present a dataset that contains over 10000 synchronization strategies and run-time pairs on a diverse set of models and cluster specifications.

Beyond local convolution networks, we explore how to harness various external human knowledge for endowing the networks with the capability of semantic global reasoning. Rather than using separate graphical models (e.g. CRF) or constraints for modeling broader dependencies, we propose a new Symbolic Graph Reasoning (SGR) layer, which performs reasoning over a group of symbolic nodes whose outputs explicitly represent different properties of each semantic in a prior knowledge graph. To cooperate with local convolutions, each SGR is constituted by three modules: a) a primal local-to-semantic voting module where the features of all symbolic nodes are generated by voting from local representations; b) a graph reasoning module propagates information over knowledge graph to achieve global semantic coherency; c) a dual semantic-to-local mapping module learns new associations of the evolved symbolic nodes with local representations, and accordingly enhances local features. The SGR layer can be injected between any convolution layers and instantiated with distinct prior graphs. Extensive experiments show incorporating SGR significantly improves plain ConvNets on three semantic segmentation tasks and one image classification task. More analyses show the SGR layer learns shared symbolic representations for domains/datasets with the different label set given a universal knowledge graph, demonstrating its superior generalization capability.

A natural image usually conveys rich semantic content and can be viewed from different angles. Existing image description methods are largely restricted by small sets of biased visual paragraph annotations, and fail to cover rich underlying semantics. In this paper, we investigate a semi-supervised paragraph generative framework that is able to synthesize diverse and semantically coherent paragraph descriptions by reasoning over local semantic regions and exploiting linguistic knowledge. The proposed Recurrent Topic-Transition Generative Adversarial Network (RTT-GAN) builds an adversarial framework between a structured paragraph generator and multi-level paragraph discriminators. The paragraph generator generates sentences recurrently by incorporating region-based visual and language attention mechanisms at each step. The quality of generated paragraph sentences is assessed by multi-level adversarial discriminators from two aspects, namely, plausibility at sentence level and topic-transition coherence at paragraph level. The joint adversarial training of RTT-GAN drives the model to generate realistic paragraphs with smooth logical transition between sentence topics. Extensive quantitative experiments on image and video paragraph datasets demonstrate the effectiveness of our RTT-GAN in both supervised and semi-supervised settings. Qualitative results on telling diverse stories for an image verify the interpretability of RTT-GAN.

We study the problem of conditional generative modeling based on designated semantics or structures. Existing models that build conditional generators either require massive labeled instances as supervision or are unable to accurately control the semantics of generated samples. We propose structured generative adversarial networks (SGANs) for semi-supervised conditional generative modeling. SGAN assumes the data x is generated conditioned on two independent latent variables: y that encodes the designated semantics, and z that contains other factors of variation. To ensure disentangled semantics in y and z, SGAN builds two collaborative games in the hidden space to minimize the reconstruction error of y and z, respectively. Training SGAN also involves solving two adversarial games that have their equilibrium concentrating at the true joint data distributions p(x, z) and p(x, y), avoiding distributing the probability mass diffusely over data space that MLE-based methods may suffer. We assess SGAN by evaluating its trained networks, and its performance on downstream tasks. We show that SGAN delivers a highly controllable generator, and disentangled representations; it also establishes start-of-the-art results across multiple datasets when applied for semi-supervised image classification (1.27%, 5.73%, 17.26% error rates on MNIST, SVHN and CIFAR-10 using 50, 1000 and 4000 labels, respectively). Benefiting from the separate modeling of y and z, SGAN can generate images with high visual quality and strictly following the designated semantic, and can be extended to a wide spectrum of applications, such as style transfer.

We study the problem of automatically building hypernym taxonomies from textual and visual data.Previous works in taxonomy induction generally ignore the increasingly prominent visual data, which encode important perceptual semantics.Instead, we propose a probabilistic model for taxonomy induction by jointly leveraging text and images.To avoid hand-crafted feature engineering, we design end-to-end features based on distributed representations of images and words.The model is discriminatively trained given a small set of existing ontologies and is capable of building full taxonomies from scratch for a collection of unseen conceptual label items with associated images.We evaluate our model and features on the WordNet hierarchies, where our system outperforms previous approaches by a large gap.

Large-scale deep learning requires huge computational resources to train a multi-layer neural network. Recent systems propose using 100s to 1000s of machines to train networks with tens of layers and billions of connections. While the computation involved can be done more efficiently on GPUs than on more traditional CPU cores, training such networks on a single GPU is too slow and training on distributed GPUs can be inefficient, due to data movement overheads, GPU stalls, and limited GPU memory. This paper describes a new parameter server, called GeePS, that supports scalable deep learning across GPUs distributed among multiple machines, overcoming these obstacles. We show that GeePS enables a state-of-the-art single-node GPU implementation to scale well, such as to 13 times the number of training images processed per second on 16 machines (relative to the original optimized single-node code). Moreover, GeePS achieves a higher training throughput with just four GPU machines than that a state-of-the-art CPU-only system achieves with 108 machines.

Photo retouching enables photographers to invoke dramatic visual impressions by artistically enhancing their photos through stylistic color and tone adjustments. However, it is also a time-consuming and challenging task that requires advanced skills beyond the abilities of casual photographers. Using an automated algorithm is an appealing alternative to manual work, but such an algorithm faces many hurdles. Many photographic styles rely on subtle adjustments that depend on the image content and even its semantics. Further, these adjustments are often spatially varying. Existing automatic algorithms are still limited and cover only a subset of these challenges. Recently, deep learning has shown unique abilities to address hard problems. This motivated us to explore the use of deep neural networks (DNNs) in the context of photo editing. In this article, we formulate automatic photo adjustment in a manner suitable for this approach. We also introduce an image descriptor accounting for the local semantics of an image. Our experiments demonstrate that training DNNs using these descriptors successfully capture sophisticated photographic styles. In particular and unlike previous techniques, it can model local adjustments that depend on image semantics. We show that this yields results that are qualitatively and quantitatively better than previous work.

In particular for “low resource” Keyword Search (KWS) and Speech-to-Text (STT) tasks, more untranscribed test data may be available than training data. Several approaches have been proposed to make this data useful during system development, even when initial systems have Word Error Rates (WER) above 70%. In this paper, we present a set of experiments on low-resource languages in telephony speech quality in Assamese, Bengali, Lao, Haitian, Zulu, and Tamil, demonstrating the impact that such techniques can have, in particular learning robust bottle-neck features on the test data. In the case of Tamil, when significantly more test data than training data is available, we integrated semi-supervised training and speaker adaptation on the test data, and achieved significant additional improvements in STT and KWS.

Dropout and DropConnect can be viewed as regularization methods for deep neural network (DNN) training. In DNN acoustic modeling, the huge number of speech samples makes it expensive to sample the neuron mask (Dropout) or the weight mask (DropConnect) repetitively from a high dimensional distribution. In this paper we investigate the effect of Gaussian stochastic neurons on DNN acoustic modeling. The pre-Gaussian stochastic term can be viewed as a variant of Dropout/DropConnect and the post-Gaussian stochastic term generalizes the idea of data augmentation into hidden layers. Gaussian stochastic neurons can give improvement on large data sets where Dropout tends to be less useful. Under the low resource condition, its performance is comparable with Dropout, but with a lower time complexity during fine-tuning.

In image classification, visual separability between different object categories is highly uneven, and some categories are more difficult to distinguish than others. Such difficult categories demand more dedicated classifiers. However, existing deep convolutional neural networks (CNN) are trained as flat N-way classifiers, and few efforts have been made to leverage the hierarchical structure of categories. In this paper, we introduce hierarchical deep CNNs (HD-CNNs) by embedding deep CNNs into a two-level category hierarchy. An HD-CNN separates easy classes using a coarse category classifier while distinguishing difficult classes using fine category classifiers. During HDCNN training, component-wise pretraining is followed by global fine-tuning with a multinomial logistic loss regularized by a coarse category consistency term. In addition, conditional executions of fine category classifiers and layer parameter compression make HD-CNNs scalable for largescale visual recognition. We achieve state-of-the-art results on both CIFAR100 and large-scale ImageNet 1000-class benchmark datasets. In our experiments, we build up three different two-level HD-CNNs, and they lower the top-1 error of the standard CNNs by 2:65%, 3:1%, and 1:1%.

We propose a dynamic topic model for monitoring temporal evolution of market competition by jointly leveraging tweets and their associated images. For a market of interest (e.g. luxury goods), we aim at automatically detecting the latent topics (e.g. bags, clothes, luxurious) that are competitively shared by multiple brands (e.g. Burberry, Prada, and Chanel), and tracking temporal evolution of the brands' stakes over the shared topics. One of key applications of our work is social media monitoring that can provide companies with temporal summaries of highly overlapped or discriminative topics with their major competitors. We design our model to correctly address three major challenges: multiview representation of text and images, modeling of competitiveness of multiple brands over shared topics, and tracking their temporal evolution. As far as we know, no previous model can satisfy all the three challenges. For evaluation, we analyze about 10 millions of tweets and 8 millions of associated images of the 23 brands in the two categories of luxury and beer. Through experiments, we show that the proposed approach is more successful than other candidate methods for the topic modeling of competition. We also quantitatively demonstrate the generalization power of the proposed method for three prediction tasks.

In acoustic modeling, speaker adaptive training (SAT) has been a long-standing technique for the traditional Gaussian mixture models (GMMs). Acoustic models trained with SAT become independent of training speakers and generalize better to unseen testing speakers. This paper ports the idea of SAT to deep neural networks (DNNs), and proposes a framework to perform feature-space SAT for DNNs. Using i-vectors as speaker representations, our framework learns an adaptation neural network to derive speaker-normalized features. Speaker adaptive models are obtained by fine-tuning DNNs in such a feature space. This framework can be applied to various feature types and network structures, posing a very general SAT solution. In this paper, we fully investigate how to build SAT-DNN models effectively and efficiently. First, we study the optimal configurations of SAT-DNNs for large-scale acoustic modeling tasks. Then, after presenting detailed comparisons between SAT-DNNs and the existing DNN adaptation methods, we propose to combine SAT-DNNs and model-space DNN adaptation during decoding. Finally, to accelerate learning of SAT-DNNs, a simple yet effective strategy, frame skipping, is employed to reduce the size of training data. Our experiments show that compared with a strong DNN baseline, the SAT-DNN model achieves 13.5% and 17.5% relative improvement on word error rates (WERs), without and with model-space adaptation applied respectively. Data reduction based on frame skipping results in 2 × speed-up for SAT-DNN training, while causing negligible WER loss on the testing data.

Multilingual deep neural networks (DNNs) can act as deep feature extractors and have been applied successfully to crosslanguage acoustic modeling. Learning these feature extractors becomes an expensive task, because of the enlarged multilingual training data and the sequential nature of stochastic gradient descent (SGD). This paper investigates strategies to accelerate the learning process over multiple GPU cards. We propose the DistModel and DistLang frameworks which distribute feature extractor learning by models and languages respectively. The time-synchronous DistModel has the nice property of tolerating infrequent model averaging. With 3 GPUs, DistModel achieves 2.6× speed-up and causes no loss on word error rates. When using DistLang, we observe better acceleration but worse recognition performance. Further evaluations are conducted to scale DistModel to more languages and GPU cards.

We investigate the concept of speaker adaptive training (SAT) in the context of deep neural network (DNN) acoustic models. Previous studies have shown success of performing speaker adaptation for DNNs in speech recognition. In this paper, we apply SAT to DNNs by learning two types of feature mapping neural networks. Given an initial DNN model, these networks take speaker i-vectors as additional information and project DNN inputs into a speaker-normalized space. The final SAT model is obtained by updating the canonical DNN in the normalized feature space. Experiments on a Switchboard 110- hour setup show that compared with the baseline DNN, the SAT-DNN model brings 7.5% and 6.0% relative improvement when DNN inputs are speaker-independent and speakeradapted features respectively. Further evaluations on the more challenging BABEL datasets reveal significant word error rate reduction achieved by SAT-DNN.

Speaker adaptive training (SAT) is a well studied technique for Gaussian mixture acoustic models (GMMs). Recently we proposed to perform SAT for deep neural networks (DNNs), with speaker i-vectors applied in feature learning. The resulting SAT-DNN models significantly outperform DNNs on word error rates (WERs). In this paper, we present different methods to further improve and extend SAT-DNN. First, we conduct detailed analysis to investigate i-vector extractor training and flexible feature fusion. Second, the SAT-DNN approach is extended to improve tasks including bottleneck feature (BNF) generation, convolutional neural network (CNN) acoustic modeling and multilingual DNN-based feature extraction. Third, for transcribing multimedia data, we enrich the i-vector representation with global speaker attributes (age, gender, etc.) obtained automatically from video signals. On a collection of instructional videos, incorporation of the additional visual features is observed to boost the recognition accuracy of SAT-DNN.