Training deep networks requires various design decisions regarding for instance their architecture, data augmentation, or optimization. In this work, we find these training variations to result in networks learning unique feature sets from the data. Using public model libraries comprising thousands of models trained on canonical datasets like ImageNet, we observe that for arbitrary pairings of pretrained models, one model extracts significant data context unavailable in the other – independent of overall performance. Given any arbitrary pairing of pretrained models and no external rankings (such as separate test sets, e.g. due to data privacy), we investigate if it is possible to transfer such "complementary" knowledge from one model to another without performance degradation – a task made particularly difficult as additional knowledge can be contained in stronger, equiperformant or weaker models. Yet facilitating robust transfer in scenarios agnostic to pretrained model pairings would unlock auxiliary gains and knowledge fusion from any model repository without restrictions on model and problem specifics - including from weaker, lower-performance models. This work therefore provides an initial, in-depth exploration on the viability of such general-purpose knowledge transfer. Across large-scale experiments, we first reveal the shortcomings of standard knowledge distillation techniques, and then propose a much more general extension through data partitioning for successful transfer between nearly all pretrained models, which we show can also be done unsupervised. Finally, we assess both the scalability and impact of fundamental model properties on successful model-agnostic knowledge transfer.

Multimodal foundation models serve numerous applications at the intersection of vision and language. Still, despite being pretrained on extensive data, they become outdated over time.To keep models updated, research into continual pretraining mainly explores scenarios with either (1) infrequent, indiscriminate updates on large-scale new data, or (2) frequent, sample-level updates.However, practical model deployment often operates in the gap between these two limit cases, as real-world applications demand adaptation to specific subdomains, tasks or concepts --- spread over the entire, varying life cycle of a model. In this work, we complement current perspectives on continual pretraining through a research test bed and offer comprehensive guidance for effective continual model updates in such scenarios.We first introduce FoMo-in-Flux, a continual multimodal pretraining benchmark with realistic compute constraints and practical deployment requirements, constructed over 63 datasets with diverse visual and semantic coverage.Using FoMo-in-Flux, we explore the complex landscape of practical continual pretraining through multiple perspectives: (1) data mixtures and stream orderings that emulate real-world deployment settings, (2) methods ranging from simple fine-tuning and traditional continual learning strategies to parameter-efficient updates and model merging, (3) meta-learning-rate schedules and mechanistic design choices, and (4) model and compute scaling. Together, our insights provide a practitioner's guide to continual multimodal pretraining for real-world deployment. Benchmark and code is provided here: https://github.com/ExplainableML/fomoinflux.

The visual classification performance of vision-language models such as CLIP has been shown to benefit from additional semantic knowledge from large language models (LLMs) such as GPT-3. In particular, averaging over LLM-generated class descriptors, e.g. "waffle, which has a round shape", can notably improve generalization performance. In this work, we critically study this behavior and propose WaffleCLIP, a framework for zero-shot visual classification which simply replaces LLM-generated descriptors with random character and word descriptors. Without querying external models, we achieve comparable performance gains on a large number of visual classification tasks. This allows WaffleCLIP to both serve as a low-cost alternative, as well as a sanity check for any future LLM-based vision-language model extensions. We conduct an extensive experimental study on the impact and shortcomings of additional semantics introduced with LLM-generated descriptors, and showcase how - if available - semantic context is better leveraged by querying LLMs for high-level concepts, which we show can be done to jointly resolve potential class name ambiguities. Code is available here: https://github.com/ExplainableML/WaffleCLIP.

Deep Metric Learning (DML) aims to learn representation spaces on which semantic relations can simply be expressed through predefined distance metrics. Best performing approaches commonly leverage class proxies as sample stand-ins for better convergence and generalization. However, these proxy-methods solely optimize for sample-proxy distances. Given the inherent non-bijectiveness of used distance functions, this can induce locally isotropic sample distributions, leading to crucial semantic context being missed due to difficulties resolving local structures and intraclass relations between samples. To alleviate this problem, we propose non-isotropy regularization$(\mathbb{NIR})$for proxy-based Deep Metric Learning. By leveraging Normalizing Flows, we enforce unique translatability of samples from their respective class proxies. This allows us to explicitly induce a non-isotropic distribution of samples around a proxy to optimize for. In doing so, we equip proxy-based objectives to better learn local structures. Extensive experiments highlight consistent generalization benefits of NIR while achieving competitive and state-of-the-art performance on the standard benchmarks CUB200-2011, Cars196 and Stanford Online Products. In addition, we find the superior convergence properties of proxy-based methods to still be retained or even improved, making NIR very attractive for practical usage. Code available at github.com/ExplainableML/NonIsotropicProxyDML.

Deep Metric Learning (DML) proposes to learn metric spaces which encode semantic similarities as embedding space distances. These spaces should be transferable to classes beyond those seen during training. Commonly, DML methods task networks to solve contrastive ranking tasks defined over binary class assignments. However, such approaches ignore higher-level semantic relations between the actual classes. This causes learned embedding spaces to encode incomplete semantic context and misrepresent the semantic relation between classes, impacting the generalizability of the learned metric space. To tackle this issue, we propose a language guidance objective for visual similarity learning. Leveraging language embeddings of expert- and pseudo-classnames, we contextualize and realign visual representation spaces corresponding to meaningful language semantics for better semantic consistency. Extensive experiments and ablations provide a strong motivation for our proposed approach and show language guidance offering significant, model-agnostic improvements for DML, achieving competitive and state-of-the-art results on all benchmarks. Code available at github.com/ExplainableML/LanguageGuidance-for_DML.

We enhance auto-regressive language models by conditioning on document chunks retrieved from a large corpus, based on local similarity with preceding tokens. With a 2 trillion token database, our Retrieval-Enhanced Transformer (RETRO) obtains comparable performance to GPT-3 and Jurassic-1 on the Pile, despite using 25{\texttimes} fewer parameters. After fine-tuning, RETRO performance translates to downstream knowledge-intensive tasks such as question answering. RETRO combines a frozen Bert retriever, a differentiable encoder and a chunked cross-attention mechanism to predict tokens based on an order of magnitude more data than what is typically consumed during training. We typically train RETRO from scratch, yet can also rapidly RETROfit pre-trained transformers with retrieval and still achieve good performance. Our work opens up new avenues for improving language models through explicit memory at unprecedented scale.

The performance of a language model has been shown to be effectively modeled as a power-law in its parameter count. Here we study the scaling behaviors of Routing Networks: architectures that conditionally use only a subset of their parameters while processing an input. For these models, parameter count and computational requirement form two independent axes along which an increase leads to better performance. In this work we derive and justify scaling laws defined on these two variables which generalize those known for standard language models and describe the performance of a wide range of routing architectures trained via three different techniques. Afterwards we provide two applications of these laws: first deriving an Effective Parameter Count along which all models scale at the same rate, and then using the scaling coefficients to give a quantitative comparison of the three routing techniques considered. Our analysis derives from an extensive evaluation of Routing Networks across five orders of magnitude of size, including models with hundreds of experts and hundreds of billions of parameters.

Real-world data is high-dimensional: a book, image, or musical performance can easily contain hundreds of thousands of elements even after compression. However, the most commonly used autoregressive models, Transformers, are prohibitively expensive to scale to the number of inputs and layers needed to capture this long-range structure. We develop Perceiver AR, an autoregressive, modality-agnostic architecture which uses cross-attention to map long-range inputs to a small number of latents while also maintaining end-to-end causal masking. Perceiver AR can directly attend to over a hundred thousand tokens, enabling practical long-context density estimation without the need for hand-crafted sparsity patterns or memory mechanisms. When trained on images or music, Perceiver AR generates outputs with clear long-term coherence and structure. Our architecture also obtains state-of-the-art likelihood on long-sequence benchmarks, including 64x64 ImageNet images and PG-19 books.

Building models that can be rapidly adapted to novel tasks using only a handful of annotated examples is an open challenge for multimodal machine learning research. We introduce Flamingo, a family of Visual Language Models (VLM) with this ability. We propose key architectural innovations to: (i) bridge powerful pretrained vision-only and language-only models, (ii) handle sequences of arbitrarily interleaved visual and textual data, and (iii) seamlessly ingest images or videos as inputs. Thanks to their flexibility, Flamingo models can be trained on large-scale multimodal web corpora containing arbitrarily interleaved text and images, which is key to endow them with in-context few-shot learning capabilities. We perform a thorough evaluation of our models, exploring and measuring their ability to rapidly adapt to a variety of image and video tasks. These include open-ended tasks such as visual question-answering, where the model is prompted with a question which it has to answer, captioning tasks, which evaluate the ability to describe a scene or an event, and close-ended tasks such as multiple-choice visual question-answering. For tasks lying anywhere on this spectrum, a single Flamingo model can achieve a new state of the art with few-shot learning, simply by prompting the model with task-specific examples. On numerous benchmarks, Flamingo outperforms models fine-tuned on thousands of times more task-specific data.

We investigate the optimal model size and number of tokens for training a transformer language model under a given compute budget. We find that current large language models are significantly undertrained, a consequence of the recent focus on scaling language models whilst keeping the amount of training data constant. By training over 400 language models ranging from 70 million to over 16 billion parameters on 5 to 500 billion tokens, we find that for compute-optimal training, the model size and the number of training tokens should be scaled equally: for every doubling of model size the number of training tokens should also be doubled. We test this hypothesis by training a predicted compute-optimal model, Chinchilla, that uses the same compute budget as Gopher but with 70B parameters and 4$\times$ more data. Chinchilla uniformly and significantly outperformsGopher (280B), GPT-3 (175B), Jurassic-1 (178B), and Megatron-Turing NLG (530B) on a large range of downstream evaluation tasks. This also means that Chinchilla uses substantially less compute for fine-tuning and inference, greatly facilitating downstream usage. As a highlight, Chinchilla reaches a state-of-the-art average accuracy of 67.5% on the MMLU benchmark, a 7% improvement over Gopher.

The goal of this work is to recognise phrases and sentences being spoken by a talking face, with or without the audio. Unlike previous works that have focussed on recognising a limited number of words or phrases, we tackle lip reading as an open-world problem - unconstrained natural language sentences, and in the wild videos. Our key contributions are: (1) we compare two models for lip reading, one using a CTC loss, and the other using a sequence-to-sequence loss. Both models are built on top of the transformer self-attention architecture; (2) we investigate to what extent lip reading is complementary to audio speech recognition, especially when the audio signal is noisy; (3) we introduce and publicly release a new dataset for audio-visual speech recognition, LRS2-BBC, consisting of thousands of natural sentences from British television. The models that we train surpass the performance of all previous work on a lip reading benchmark dataset by a significant margin.

Over the past decade, deep learning has had a revolutionary impact on a broad range of fields such as computer vision and image processing, computational photography, medical imaging and speech and language analysis and synthesis etc. Deep learning technologies are estimated to have added billions in business value, created new markets, and transformed entire industrial segments. Most of today’s successful deep learning methods such as Convolutional Neural Networks (CNNs) rely on classical signal processing models that limit their applicability to data with underlying Euclidean grid-like structure, e.g., images or acoustic signals. Yet, many applications deal with non-Euclidean (graph- or manifold-structured) data. For example, in social network analysis the users and their attributes are generally modeled as signals on the vertices of graphs. In biology protein-to-protein interactions are modeled as graphs. In computer vision & graphics 3D objects are modeled as meshes or point clouds. Furthermore, a graph representation is a very natural way to describe interactions between objects or signals. The classical deep learning paradigm on Euclidean domains falls short in providing appropriate tools for such kind of data. Until recently, the lack of deep learning models capable of correctly dealing with non-Euclidean data has been a major obstacle in these fields. This special section addresses the need to bring together leading efforts in non-Euclidean deep learning across all communities. From the papers that the special received twelve were selected for publication. The selected papers can naturally fall in three distinct categories: (a) methodologies that advance machine learning on data that are represented as graphs, (b) methodologies that advance machine learning on manifold-valued data, and (c) applications of machine learning methodologies on non-Euclidean spaces in computer vision and medical imaging. We briefly review the accepted papers in each of the groups.

Rémi Leblond, Jean-Baptiste Alayrac, Laurent Sifre, Miruna Pislar, Lespiau Jean-Baptiste, Ioannis Antonoglou, Karen Simonyan, Oriol Vinyals. Proceedings of the 2021 Conference on Empirical Methods in Natural Language Processing. 2021.

Self-supervised pretraining has been shown to yield powerful representations for transfer learning. These performance gains come at a large computational cost however, with state-of-the-art methods requiring an order of magnitude more computation than supervised pretraining. We tackle this computational bottleneck by introducing a new self-supervised objective, contrastive detection, which tasks representations with identifying object-level features across augmentations. This objective extracts a rich learning signal per image, leading to state-of-the-art transfer accuracy on a variety of downstream tasks, while requiring up to 10× less pretraining. In particular, our strongest ImageNet-pretrained model performs on par with SEER, one of the largest self-supervised systems to date, which uses 1000× more pretraining data. Finally, our objective seamlessly handles pretraining on more complex images such as those in COCO, closing the gap with supervised transfer learning from COCO to PASCAL.

Biological systems understand the world by simultaneously processing high-dimensional inputs from modalities as diverse as vision, audition, touch, proprioception, etc. The perception models used in deep learning on the other hand are designed for individual modalities, often relying on domain-specific assumptions such as the local grid structures exploited by virtually all existing vision models. These priors introduce helpful inductive biases, but also lock models to individual modalities. In this paper we introduce the Perceiver {–} a model that builds upon Transformers and hence makes few architectural assumptions about the relationship between its inputs, but that also scales to hundreds of thousands of inputs, like ConvNets. The model leverages an asymmetric attention mechanism to iteratively distill inputs into a tight latent bottleneck, allowing it to scale to handle very large inputs. We show that this architecture is competitive with or outperforms strong, specialized models on classification tasks across various modalities: images, point clouds, audio, video and video+audio. The Perceiver obtains performance comparable to ResNet-50 and ViT on ImageNet without 2D convolutions by directly attending to 50,000 pixels. It is also competitive in all modalities in AudioSet.

Recent developments in the field of model-based RL have proven successful in a range of environments, especially ones where planning is essential. However, such successes have been limited to deterministic fully-observed environments. We present a new approach that handles stochastic and partially-observable environments. Our key insight is to use discrete autoencoders to capture the multiple possible effects of an action in a stochastic environment. We use a stochastic variant of Monte Carlo tree search to plan over both the agent’s actions and the discrete latent variables representing the environment’s response. Our approach significantly outperforms an offline version of MuZero on a stochastic interpretation of chess where the opponent is considered part of the environment. We also show that our approach scales to DeepMind Lab, a first-person 3D environment with large visual observations and partial observability.

When trained at sufficient scale, auto-regressive language models exhibit the notable ability to learn a new language task after being prompted with just a few examples. Here, we present a simple, yet effective, approach for transferring this few-shot learning ability to a multimodal setting (vision and language). Using aligned image and caption data, we train a vision encoder to represent each image as a sequence of continuous embeddings, such that a pre-trained, frozen language model presented with this prefix generates the appropriate caption. The resulting system is a multimodal few-shot learner, with the surprising ability to learn a variety of new tasks when conditioned on examples, represented as a sequence of any number of interleaved image and text embeddings. We demonstrate that it can rapidly learn words for new objects and novel visual categories, do visual question-answering with only a handful of examples, and make use of outside knowledge, by measuring a single model on a variety of established and new benchmarks.

Graph neural networks (GNNs) are typically applied to static graphs that are assumed to be known upfront. This static input structure is often informed purely by insight of the machine learning practitioner, and might not be optimal for the actual task the GNN is solving. In absence of reliable domain expertise, one might resort to inferring the latent graph structure, which is often difficult due to the vast search space of possible graphs. Here we introduce Pointer Graph Networks (PGNs) which augment sets or graphs with additional inferred edges for improved model generalisation ability. PGNs allow each node to dynamically point to another node, followed by message passing over these pointers. The sparsity of this adaptable graph structure makes learning tractable while still being sufficiently expressive to simulate complex algorithms. Critically, the pointing mechanism is directly supervised to model long-term sequences of operations on classical data structures, incorporating useful structural inductive biases from theoretical computer science. Qualitatively, we demonstrate that PGNs can learn parallelisable variants of pointer-based data structures, namely disjoint set unions and link/cut trees. PGNs generalise out-of-distribution to 5x larger test inputs on dynamic graph connectivity tasks, outperforming unrestricted GNNs and Deep Sets.

In order to efficiently transmit and store speech signals, speech codecs create a minimally redundant representation of the input signal which is then decoded at the receiver with the best possible perceptual quality. In this work we demonstrate that a neural network architecture based on VQ-VAE with a WaveNet decoder can be used to perform very low bit-rate speech coding with high reconstruction quality. A prosody-transparent and speaker-independent model trained on the LibriSpeech corpus coding audio at 1.6 kbps exhibits perceptual quality which is around halfway between the MELP codec at 2.4 kbps and AMR-WB codec at 23.05 kbps. In addition, when training on high-quality recorded speech with the test speaker included in the training set, a model coding speech at 1.6 kbps produces output of similar perceptual quality to that generated by AMR-WB at 23.05 kbps.

This paper addresses the challenging problem of retrieval and matching of graph structured objects, and makes two key contributions. First, we demonstrate how Graph Neural Networks (GNN), which have emerged as an effective model for various supervised prediction problems defined on structured data, can be trained to produce embedding of graphs in vector spaces that enables efficient similarity reasoning. Second, we propose a novel Graph Matching Network model that, given a pair of graphs as input, computes a similarity score between them by jointly reasoning on the pair through a new cross-graph attention-based matching mechanism. We demonstrate the effectiveness of our models on different domains including the challenging problem of control-flow graph based function similarity search that plays an important role in the detection of vulnerabilities in software systems. The experimental analysis demonstrates that our models are not only able to exploit structure in the context of similarity learning but they can also outperform domain specific baseline systems that have been carefully hand-engineered for these problems.

Deep generative models (DGMs) of images are now sufficiently mature that they produce nearly photorealistic samples and obtain scores similar to the data distribution on heuristics such as Frechet Inception Distance (FID). These results, especially on large-scale datasets such as ImageNet, suggest that DGMs are learning the data distribution in a perceptually meaningful space and can be used in downstream tasks. To test this latter hypothesis, we use class-conditional generative models from a number of model classes—variational autoencoders, autoregressive models, and generative adversarial networks (GANs)—to infer the class labels of real data. We perform this inference by training an image classifier using only synthetic data and using the classifier to predict labels on real data. The performance on this task, which we call Classification Accuracy Score (CAS), reveals some surprising results not identified by traditional metrics and constitute our contributions. First, when using a state-of-the-art GAN (BigGAN-deep), Top-1 and Top-5 accuracy decrease by 27.9% and 41.6%, respectively, compared to the original data; and conditional generative models from other model classes, such as Vector-Quantized Variational Autoencoder-2 (VQ-VAE-2) and Hierarchical Autoregressive Models (HAMs), substantially outperform GANs on this benchmark. Second, CAS automatically surfaces particular classes for which generative models failed to capture the data distribution, and were previously unknown in the literature. Third, we find traditional GAN metrics such as Inception Score (IS) and FID neither predictive of CAS nor useful when evaluating non-GAN models. Furthermore, in order to facilitate better diagnoses of generative models, we open-source the proposed metric.

We explore the use of Vector Quantized Variational AutoEncoder (VQ-VAE) models for large scale image generation. To this end, we scale and enhance the autoregressive priors used in VQ-VAE to generate synthetic samples of much higher coherence and fidelity than possible before. We use simple feed-forward encoder and decoder networks, making our model an attractive candidate for applications where the encoding and/or decoding speed is critical. Additionally, VQ-VAE requires sampling an autoregressive model only in the compressed latent space, which is an order of magnitude faster than sampling in the pixel space, especially for large images. We demonstrate that a multi-scale hierarchical organization of VQ-VAE, augmented with powerful priors over the latent codes, is able to generate samples with quality that rivals that of state of the art Generative Adversarial Networks on multifaceted datasets such as ImageNet, while not suffering from GAN's known shortcomings such as mode collapse and lack of diversity.

Voice activity detection (VAD) is the task of predicting which parts of an utterance contains speech versus background noise. It is an important first step to determine which samples to send to the decoder and when to close the microphone. The long short-term memory neural network (LSTM) is a popular architecture for sequential modeling of acoustic signals, and has been successfully used in several VAD applications. However, it has been observed that LSTMs suffer from state saturation problems when the utterance is long (i.e., for voice dictation tasks), and thus requires the LSTM state to be periodically reset. In this paper, we propose an alternative architecture that does not suffer from saturation problems by modeling temporal variations through a stateless dilated convolution neural network (CNN). The proposed architecture differs from conventional CNNs in three respects: it uses dilated causal convolution, gated activations and residual connections. Results on a Google Voice Typing task shows that the proposed architecture achieves 14% relative FA improvement at a FR of 1% over state-of-the-art LSTMs for VAD task. We also include detailed experiments investigating the factors that distinguish the proposed architecture from conventional convolution.

Advances in deep generative networks have led to impressive results in recent years. Nevertheless, such models can often waste their capacity on the minutiae of datasets, presumably due to weak inductive biases in their decoders. This is where graphics engines may come in handy since they abstract away low-level details and represent images as high-level programs. Current methods that combine deep learning and renderers are limited by hand-crafted likelihood or distance functions, a need for large amounts of supervision, or difficulties in scaling their inference algorithms to richer datasets. To mitigate these issues, we present SPIRAL, an adversarially trained agent that generates a program which is executed by a graphics engine to interpret and sample images. The goal of this agent is to fool a discriminator network that distinguishes between real and rendered data, trained with a distributed reinforcement learning setup without any supervision. A surprising finding is that using the discriminator’s output as a reward signal is the key to allow the agent to make meaningful progress at matching the desired output rendering. To the best of our knowledge, this is the first demonstration of an end-to-end, unsupervised and adversarial inverse graphics agent on challenging real world (MNIST, Omniglot, CelebA) and synthetic 3D datasets. A video of the agent can be found at https://youtu.be/iSyvwAwa7vk.

Planning problems are among the most important and well-studied problems in artificial intelligence. They are most typically solved by tree search algorithms that simulate ahead into the future, evaluate future states, and back-up those evaluations to the root of a search tree. Among these algorithms, Monte-Carlo tree search (MCTS) is one of the most general, powerful and widely used. A typical implementation of MCTS uses cleverly designed rules, optimised to the particular characteristics of the domain. These rules control where the simulation traverses, what to evaluate in the states that are reached, and how to back-up those evaluations. In this paper we instead learn where, what and how to search. Our architecture, which we call an MCTSnet, incorporates simulation-based search inside a neural network, by expanding, evaluating and backing-up a vector embedding. The parameters of the network are trained end-to-end using gradient-based optimisation. When applied to small searches in the well-known planning problem Sokoban, the learned search algorithm significantly outperformed MCTS baselines.

The recently-developed WaveNet architecture is the current state of the art in realistic speech synthesis, consistently rated as more natural sounding for many different languages than any previous system. However, because WaveNet relies on sequential generation of one audio sample at a time, it is poorly suited to today’s massively parallel computers, and therefore hard to deploy in a real-time production setting. This paper introduces Probability Density Distillation, a new method for training a parallel feed-forward network from a trained WaveNet with no significant difference in quality. The resulting system is capable of generating high-fidelity speech samples at more than 20 times faster than real-time, a 1000x speed up relative to the original WaveNet, and capable of serving multiple English and Japanese voices in a production setting.

We propose a method of moments (MoM) algorithm for training large-scale implicit generative models. Moment estimation in this setting encounters two problems: it is often difficult to define the millions of moments needed to learn the model parameters, and it is hard to determine which properties are useful when specifying moments. To address the first issue, we introduce a moment network, and define the moments as the network’s hidden units and the gradient of the network’s output with respect to its parameters. To tackle the second problem, we use asymptotic theory to highlight desiderata for moments – namely they should minimize the asymptotic variance of estimated model parameters – and introduce an objective to learn better moments. The sequence of objectives created by this Method of Learned Moments (MoLM) can train high-quality neural image samplers. On CIFAR-10, we demonstrate that MoLM-trained generators achieve significantly higher Inception Scores and lower Frechet Inception Distances than those trained with gradient penalty-regularized and spectrally-normalized adversarial objectives. These generators also achieve nearly perfect Multi-Scale Structural Similarity Scores on CelebA, and can create high-quality samples of 128x128 images.

Memory-based neural networks model temporal data by leveraging an ability to remember information for long periods. It is unclear, however, whether they also have an ability to perform complex relational reasoning with the information they remember. Here, we first confirm our intuitions that standard memory architectures may struggle at tasks that heavily involve an understanding of the ways in which entities are connected -- i.e., tasks involving relational reasoning. We then improve upon these deficits by using a new memory module -- a Relational Memory Core (RMC) -- which employs multi-head dot product attention to allow memories to interact. Finally, we test the RMC on a suite of tasks that may profit from more capable relational reasoning across sequential information, and show large gains in RL domains (BoxWorld & Mini PacMan), program evaluation, and language modeling, achieving state-of-the-art results on the WikiText-103, Project Gutenberg, and GigaWord datasets.

The goal of this work is to recognise phrases and sentences being spoken by a talking face, with or without the audio. Unlike previous works that have focussed on recognising a limited number of words or phrases, we tackle lip reading as an open-world problem - unconstrained natural language sentences, and in the wild videos. Our key contributions are: (1) a `Watch, Listen, Attend and Spell' (WLAS) network that learns to transcribe videos of mouth motion to characters; (2) a curriculum learning strategy to accelerate training and to reduce overfitting; (3) a `Lip Reading Sentences' (LRS) dataset for visual speech recognition, consisting of over 100,000 natural sentences from British television. The WLAS model trained on the LRS dataset surpasses the performance of all previous work on standard lip reading benchmark datasets, often by a significant margin. This lip reading performance beats a professional lip reader on videos from BBC television, and we also demonstrate that if audio is available, then visual information helps to improve speech recognition performance.

When training neural networks, the use of Synthetic Gradients (SG) allows layers or modules to be trained without update locking – without waiting for a true error gradient to be backpropagated – resulting in Decoupled Neural Interfaces (DNIs). This unlocked ability of being able to update parts of a neural network asynchronously and with only local information was demonstrated to work empirically in Jaderberg et al (2016). However, there has been very little demonstration of what changes DNIs and SGs impose from a functional, representational, and learning dynamics point of view. In this paper, we study DNIs through the use of synthetic gradients on feed-forward networks to better understand their behaviour and elucidate their effect on optimisation. We show that the incorporation of SGs does not affect the representational strength of the learning system for a neural network, and prove the convergence of the learning system for linear and deep linear models. On practical problems we investigate the mechanism by which synthetic gradient estimators approximate the true loss, and, surprisingly, how that leads to drastically different layer-wise representations. Finally, we also expose the relationship of using synthetic gradients to other error approximation techniques and find a unifying language for discussion and comparison.

Supervised learning on molecules has incredible potential to be useful in chemistry, drug discovery, and materials science. Luckily, several promising and closely related neural network models invariant to molecular symmetries have already been described in the literature. These models learn a message passing algorithm and aggregation procedure to compute a function of their entire input graph. At this point, the next step is to find a particularly effective variant of this general approach and apply it to chemical prediction benchmarks until we either solve them or reach the limits of the approach. In this paper, we reformulate existing models into a single common framework we call Message Passing Neural Networks (MPNNs) and explore additional novel variations within this framework. Using MPNNs we demonstrate state of the art results on an important molecular property prediction benchmark; these results are strong enough that we believe future work should focus on datasets with larger molecules or more accurate ground truth labels.

Training directed neural networks typically requires forward-propagating data through a computation graph, followed by backpropagating error signal, to produce weight updates. All layers, or more generally, modules, of the network are therefore locked, in the sense that they must wait for the remainder of the network to execute forwards and propagate error backwards before they can be updated. In this work we break this constraint by decoupling modules by introducing a model of the future computation of the network graph. These models predict what the result of the modelled subgraph will produce using only local information. In particular we focus on modelling error gradients: by using the modelled synthetic gradient in place of true backpropagated error gradients we decouple subgraphs, and can update them independently and asynchronously i.e. we realise decoupled neural interfaces. We show results for feed-forward models, where every layer is trained asynchronously, recurrent neural networks (RNNs) where predicting one’s future gradient extends the time over which the RNN can effectively model, and also a hierarchical RNN system with ticking at different timescales. Finally, we demonstrate that in addition to predicting gradients, the same framework can be used to predict inputs, resulting in models which are decoupled in both the forward and backwards pass – amounting to independent networks which co-learn such that they can be composed into a single functioning corporation.

We propose a probabilistic video model, the Video Pixel Network (VPN), that estimates the discrete joint distribution of the raw pixel values in a video. The model and the neural architecture reflect the time, space and color structure of video tensors and encode it as a four-dimensional dependency chain. The VPN approaches the best possible performance on the Moving MNIST benchmark, a leap over the previous state of the art, and the generated videos show only minor deviations from the ground truth. The VPN also produces detailed samples on the action-conditional Robotic Pushing benchmark and generalizes to the motion of novel objects.

Deep reinforcement learning methods attain super-human performance in a wide range of environments. Such methods are grossly inefficient, often taking orders of magnitudes more data than humans to achieve reasonable performance. We propose Neural Episodic Control: a deep reinforcement learning agent that is able to rapidly assimilate new experiences and act upon them. Our agent uses a semi-tabular representation of the value function: a buffer of past experience containing slowly changing state representations and rapidly updated estimates of the value function. We show across a wide range of environments that our agent learns significantly faster than other state-of-the-art, general purpose deep reinforcement learning agents.

Learning useful representations without supervision remains a key challenge in machine learning. In this paper, we propose a simple yet powerful generative model that learns such discrete representations. Our model, the Vector Quantised-Variational AutoEncoder (VQ-VAE), differs from VAEs in two key ways: the encoder network outputs discrete, rather than continuous, codes; and the prior is learnt rather than static. In order to learn a discrete latent representation, we incorporate ideas from vector quantisation (VQ). Using the VQ method allows the model to circumvent issues of ``posterior collapse'' -— where the latents are ignored when they are paired with a powerful autoregressive decoder -— typically observed in the VAE framework. Pairing these representations with an autoregressive prior, the model can generate high quality images, videos, and speech as well as doing high quality speaker conversion and unsupervised learning of phonemes, providing further evidence of the utility of the learnt representations.

We introduce Imagination-Augmented Agents (I2As), a novel architecture for deep reinforcement learning combining model-free and model-based aspects. In contrast to most existing model-based reinforcement learning and planning methods, which prescribe how a model should be used to arrive at a policy, I2As learn to interpret predictions from a trained environment model to construct implicit plans in arbitrary ways, by using the predictions as additional context in deep policy networks. I2As show improved data efficiency, performance, and robustness to model misspecification compared to several strong baselines.

Automatically describing the content of an image is a fundamental problem in artificial intelligence that connects computer vision and natural language processing. In this paper, we present a generative model based on a deep recurrent architecture that combines recent advances in computer vision and machine translation and that can be used to generate natural sentences describing an image. The model is trained to maximize the likelihood of the target description sentence given the training image. Experiments on several datasets show the accuracy of the model and the fluency of the language it learns solely from image descriptions. Our model is often quite accurate, which we verify both qualitatively and quantitatively. Finally, given the recent surge of interest in this task, a competition was organized in 2015 using the newly released COCO dataset. We describe and analyze the various improvements we applied to our own baseline and show the resulting performance in the competition, which we won ex-aequo with a team from Microsoft Research.

The standard recurrent neural network language model (rnnlm) generates sentences one word at a time and does not work from an explicit global sentence representation.In this work, we introduce and study an rnn-based variational autoencoder generative model that incorporates distributed latent representations of entire sentences.This factorization allows it to explicitly model holistic properties of sentences such as style, topic, and high-level syntactic features.Samples from the prior over these sentence representations remarkably produce diverse and well-formed sentences through simple deterministic decoding.By examining paths through this latent space, we are able to generate coherent novel sentences that interpolate between known sentences.We present techniques for solving the difficult learning problem presented by this model, demonstrate its effectiveness in imputing missing words, explore many interesting properties of the model's latent sentence space, and present negative results on the use of the model in language modeling.but now , as they parked out front and owen stepped out of the car , he could see True: that the transition was complete .RNNLM: it , " i said .VAE: through the driver 's door .you kill

We present Listen, Attend and Spell (LAS), a neural speech recognizer that transcribes speech utterances directly to characters without pronunciation models, HMMs or other components of traditional speech recognizers. In LAS, the neural network architecture subsumes the acoustic, pronunciation and language models making it not only an end-to-end trained system but an end-to-end model. In contrast to DNN-HMM, CTC and most other models, LAS makes no independence assumptions about the probability distribution of the output character sequences given the acoustic sequence. Our system has two components: a listener and a speller. The listener is a pyramidal recurrent network encoder that accepts filter bank spectra as inputs. The speller is an attention-based recurrent network decoder that emits each character conditioned on all previous characters, and the entire acoustic sequence. On a Google voice search task, LAS achieves a WER of 14.1% without a dictionary or an external language model and 10.3% with language model rescoring over the top 32 beams. In comparison, the state-of-the-art CLDNN-HMM model achieves a WER of 8.0% on the same set.

Dan Gillick, Cliff Brunk, Oriol Vinyals, Amarnag Subramanya. Proceedings of the 2016 Conference of the North American Chapter of the Association for Computational Linguistics: Human Language Technologies. 2016.

Sequence-to-sequence models have achieved impressive results on various tasks. However, they are unsuitable for tasks that require incremental predictions to be made as more data arrives or tasks that have long input sequences and output sequences. This is because they generate an output sequence conditioned on an entire input sequence. In this paper, we present a Neural Transducer that can make incremental predictions as more input arrives, without redoing the entire computation. Unlike sequence-to-sequence models, the Neural Transducer computes the next-step distribution conditioned on the partially observed input sequence and the partially generated sequence. At each time step, the transducer can decide to emit zero to many output symbols. The data can be processed using an encoder and presented as input to the transducer. The discrete decision to emit a symbol at every time step makes it difficult to learn with conventional backpropagation. It is however possible to train the transducer by using a dynamic programming algorithm to generate target discrete decisions. Our experiments show that the Neural Transducer works well in settings where it is required to produce output predictions as data come in. We also find that the Neural Transducer performs well for long sequences even when attention mechanisms are not used.

This work explores conditional image generation with a new image density model based on the PixelCNN architecture. The model can be conditioned on any vector, including descriptive labels or tags, or latent embeddings created by other networks. When conditioned on class labels from the ImageNet database, the model is able to generate diverse, realistic scenes representing distinct animals, objects, landscapes and structures. When conditioned on an embedding produced by a convolutional network given a single image of an unseen face, it generates a variety of new portraits of the same person with different facial expressions, poses and lighting conditions. We also show that conditional PixelCNN can serve as a powerful decoder in an image autoencoder. Additionally, the gated convolutional layers in the proposed model improve the log-likelihood of PixelCNN to match the state-of-the-art performance of PixelRNN on ImageNet, with greatly reduced computational cost.

We present a novel deep recurrent neural network architecture that learns to build implicit plans in an end-to-end manner purely by interacting with an environment in reinforcement learning setting. The network builds an internal plan, which is continuously updated upon observation of the next input from the environment. It can also partition this internal representation into contiguous sub-sequences by learning for how long the plan can be committed to -- i.e. followed without replaning. Combining these properties, the proposed model, dubbed STRategic Attentive Writer (STRAW) can learn high-level, temporally abstracted macro-actions of varying lengths that are solely learnt from data without any prior information. These macro-actions enable both structured exploration and economic computation. We experimentally demonstrate that STRAW delivers strong improvements on several ATARI games by employing temporally extended planning strategies (e.g. Ms. Pacman and Frostbite). It is at the same time a general algorithm that can be applied on any sequence data. To that end, we also show that when trained on text prediction task, STRAW naturally predicts frequent n-grams (instead of macro-actions), demonstrating the generality of the approach.

Learning from a few examples remains a key challenge in machine learning. Despite recent advances in important domains such as vision and language, the standard supervised deep learning paradigm does not offer a satisfactory solution for learning new concepts rapidly from little data. In this work, we employ ideas from metric learning based on deep neural features and from recent advances that augment neural networks with external memories. Our framework learns a network that maps a small labelled support set and an unlabelled example to its label, obviating the need for fine-tuning to adapt to new class types. We then define one-shot learning problems on vision (using Omniglot, ImageNet) and language tasks. Our algorithm improves one-shot accuracy on ImageNet from 82.2% to 87.8% and from 88% accuracy to 95% accuracy on Omniglot compared to competing approaches. We also demonstrate the usefulness of the same model on language modeling by introducing a one-shot task on the Penn Treebank.

Thang Luong, Ilya Sutskever, Quoc Le, Oriol Vinyals, Wojciech Zaremba. Proceedings of the 53rd Annual Meeting of the Association for Computational Linguistics and the 7th International Joint Conference on Natural Language Processing (Volume 1: Long Papers). 2015.

Convolutional neural networks (CNNs) have been extensively applied for image recognition problems giving state-of-the-art results on recognition, detection, segmentation and retrieval. In this work we propose and evaluate several deep neural network architectures to combine image information across a video over longer time periods than previously attempted. We propose two methods capable of handling full length videos. The first method explores various convolutional temporal feature pooling architectures, examining the various design choices which need to be made when adapting a CNN for this task. The second proposed method explicitly models the video as an ordered sequence of frames. For this purpose we employ a recurrent neural network that uses Long Short-Term Memory (LSTM) cells which are connected to the output of the underlying CNN. Our best networks exhibit significant performance improvements over previously published results on the Sports 1 million dataset (73.1% vs. 60.9%) and the UCF-101 datasets with (88.6% vs. 88.0%) and without additional optical flow information (82.6% vs. 73.0%).

Automatically describing the content of an image is a fundamental problem in artificial intelligence that connects computer vision and natural language processing. In this paper, we present a generative model based on a deep recurrent architecture that combines recent advances in computer vision and machine translation and that can be used to generate natural sentences describing an image. The model is trained to maximize the likelihood of the target description sentence given the training image. Experiments on several datasets show the accuracy of the model and the fluency of the language it learns solely from image descriptions. Our model is often quite accurate, which we verify both qualitatively and quantitatively. For instance, while the current state-of-the-art BLEU-1 score (the higher the better) on the Pascal dataset is 25, our approach yields 59, to be compared to human performance around 69. We also show BLEU-1 score improvements on Flickr30k, from 56 to 66, and on SBU, from 19 to 28. Lastly, on the newly released COCO dataset, we achieve a BLEU-4 of 27.7, which is the current state-of-the-art.

We present an LSTM approach to deletion-based sentence compression where the task is to translate a sentence into a sequence of zeros and ones, corresponding to token deletion decisions.We demonstrate that even the most basic version of the system, which is given no syntactic information (no PoS or NE tags, or dependencies) or desired compression length, performs surprisingly well: around 30% of the compressions from a large test set could be regenerated.We compare the LSTM system with a competitive baseline which is trained on the same amount of data but is additionally provided with all kinds of linguistic features.In an experiment with human raters the LSTMbased model outperforms the baseline achieving 4.5 in readability and 3.8 in informativeness.

Both Convolutional Neural Networks (CNNs) and Long Short-Term Memory (LSTM) have shown improvements over Deep Neural Networks (DNNs) across a wide variety of speech recognition tasks. CNNs, LSTMs and DNNs are complementary in their modeling capabilities, as CNNs are good at reducing frequency variations, LSTMs are good at temporal modeling, and DNNs are appropriate for mapping features to a more separable space. In this paper, we take advantage of the complementarity of CNNs, LSTMs and DNNs by combining them into one unified architecture. We explore the proposed architecture, which we call CLDNN, on a variety of large vocabulary tasks, varying from 200 to 2,000 hours. We find that the CLDNN provides a 4-6% relative improvement in WER over an LSTM, the strongest of the three individual models.

Learning an acoustic model directly from the raw waveform has been an active area of research. However, waveform-based models have not yet matched the performance of log-mel trained neural networks. We will show that raw wave-form features match the performance of log-mel filterbank ener-gies when used with a state-of-the-art CLDNN acoustic model trained on over 2,000 hours of speech. Specifically, we will show the benefit of the CLDNN, namely the time convolution layer in reducing temporal variations, the frequency convolution layer for preserving locality and reducing frequency variations, as well as the LSTM layers for temporal modeling. In addition, by stacking raw waveform features with log-mel features, we achieve a 3 % relative reduction in word error rate. 1.

Recurrent Neural Networks can be trained to produce sequences of tokens given some input, as exemplified by recent results in machine translation and image captioning. The current approach to training them consists of maximizing the likelihood of each token in the sequence given the current (recurrent) state and the previous token. At inference, the unknown previous token is then replaced by a token generated by the model itself. This discrepancy between training and inference can yield errors that can accumulate quickly along the generated sequence. We propose a curriculum learning strategy to gently change the training process from a fully guided scheme using the true previous token, towards a less guided scheme which mostly uses the generated token instead. Experiments on several sequence prediction tasks show that this approach yields significant improvements. Moreover, it was used successfully in our winning bid to the MSCOCO image captioning challenge, 2015.

We introduce a new neural architecture to learn the conditional probability of an output sequence with elements that arediscrete tokens corresponding to positions in an input sequence.Such problems cannot be trivially addressed by existent approaches such as sequence-to-sequence and Neural Turing Machines,because the number of target classes in eachstep of the output depends on the length of the input, which is variable.Problems such as sorting variable sized sequences, and various combinatorialoptimization problems belong to this class. Our model solvesthe problem of variable size output dictionaries using a recently proposedmechanism of neural attention. It differs from the previous attentionattempts in that, instead of using attention to blend hidden units of anencoder to a context vector at each decoder step, it uses attention asa pointer to select a member of the input sequence as the output. We call this architecture a Pointer Net (Ptr-Net).We show Ptr-Nets can be used to learn approximate solutions to threechallenging geometric problems -- finding planar convex hulls, computingDelaunay triangulations, and the planar Travelling Salesman Problem-- using training examples alone. Ptr-Nets not only improve oversequence-to-sequence with input attention, butalso allow us to generalize to variable size output dictionaries.We show that the learnt models generalize beyond the maximum lengthsthey were trained on. We hope our results on these taskswill encourage a broader exploration of neural learning for discreteproblems.

Syntactic constituency parsing is a fundamental problem in naturallanguage processing which has been the subject of intensive researchand engineering for decades. As a result, the most accurate parsersare domain specific, complex, and inefficient. In this paper we showthat the domain agnostic attention-enhanced sequence-to-sequence modelachieves state-of-the-art results on the most widely used syntacticconstituency parsing dataset, when trained on a large synthetic corpusthat was annotated using existing parsers. It also matches theperformance of standard parsers when trained on a smallhuman-annotated dataset, which shows that this model is highlydata-efficient, in contrast to sequence-to-sequence models without theattention mechanism. Our parser is also fast, processing over ahundred sentences per second with an unoptimized CPU implementation.

In this paper we propose to directly optimize a discrete objective function by smoothing it, showing it is both effective at enhancing the figure of merit that we are interested in while keeping the overall complexity of the training procedure unaltered. We looked at the task of keyword detection with data scarcity (e.g., for languages for which we do not have enough data), and found it useful to optimize the Actual Term Weighted Value (ATWV) directly. In particular, we were able to automatically set the detection threshold while improving ATWV by more than 1% using a computationally cheap method based on a smoothed ATWV on both single systems and for system combination. Furthermore, we did study additional features to refine keyword candidates which were easy to optimize thanks to the same techniques, and improved ATWV by an additional 1%. The advantage of our method with respect to others is that, since we can use continuous optimization techniques, it does not impose a limit in the number of parameters that other discrete optimization techniques exhibit.

We evaluate whether features extracted from the activation of a deep convolutional network trained in a fully supervised fashion on a large, fixed set of object recognition tasks can be re-purposed to novel generic tasks. Our generic tasks may differ significantly from the originally trained tasks and there may be insufficient labeled or unlabeled data to conventionally train or adapt a deep architecture to the new tasks. We investigate and visualize the semantic clustering of deep convolutional features with respect to a variety of such tasks, including scene recognition, domain adaptation, and fine-grained recognition challenges. We compare the efficacy of relying on various network levels to define a fixed feature, and report novel results that significantly outperform the state-of-the-art on several important vision challenges. We are releasing DeCAF, an open-source implementation of these deep convolutional activation features, along with all associated network parameters to enable vision researchers to be able to conduct experimentation with deep representations across a range of visual concept learning paradigms.

Feature encoding with an overcomplete dictionary has demonstrated good performance in many applications, especially computer vision. In this paper we analyze the classification accuracy with respect to dictionary size by linking the encoding stage to kernel methods and \nystrom sampling, and obtain useful bounds on accuracy as a function of size. The \nystrom method also inspires us to revisit dictionary learning from local patches, and we propose to learn the dictionary in an end-to-end fashion taking into account pooling, a common computational layer in vision. We validate our contribution by showing how the derived bounds are able to explain the observed behavior of multiple datasets, and show that the pooling aware method efficiently reduces the dictionary size by a factor of two for a given accuracy.

In this paper, we show how new training principles and optimization techniques for neural networks can be used for different network structures. In particular, we revisit the Recurrent Neural Network (RNN), which explicitly models the Markovian dynamics of a set of observations through a non-linear function with a much larger hidden state space than traditional sequence models such as an HMM. We apply pretraining principles used for Deep Neural Networks (DNNs) and second-order optimization techniques to train an RNN. Moreover, we explore its application in the Aurora2 speech recognition task under mismatched noise conditions using a Tandem approach. We observe top performance on clean speech, and under high noise conditions, compared to multi-layer perceptrons (MLPs) and DNNs, with the added benefit of being a “deeper” model than an MLP but more compact than a DNN.

A key challenge in developing conversational systems is fusing streams of information provided by different sensors to make inferences about the behaviors and goals of people. Such systems can leverage visual and audio information collected through cameras and microphone arrays, including the location of various people, their focus of attention, body pose, the sound source direction, prosody, and speech recognition results. In this paper, we explore discriminative learning techniques for making accurate inferences on the problems of speaker, addressee and overlap detection in multiparty human-computer dialog. The focus is on finding ways to leverage within- and across-signal temporal patterns and to automatically construct representations from the raw streams that are informative for the inference problem. We present a novel extension to traditional decision trees which allows them to incorporate and model temporal signals. We contrast these methods with more traditional approaches where a human expert manually engineers relevant temporal features. The proposed approach performs well even with relatively small amounts of training data, which is of practical importance as designing features that are task dependent is time consuming and not always possible.

Recent work on deep neural networks as acoustic models for automatic speech recognition (ASR) have demonstrated substantial performance improvements. We introduce a model which uses a deep recurrent auto encoder neural network to denoise input features for robust ASR. The model is trained on stereo (noisy and clean) audio features to predict clean features given noisy input. The model makes no assumptions about how noise affects the signal, nor the existence of distinct noise environments. Instead, the model can learn to model any type of distortion or additive noise given sufficient training data. We demonstrate the model is competitive with existing feature denoising approaches on the Aurora2 task, and outperforms a tandem approach where deep networks are used to predict phoneme posteriors directly. Index Terms: neural networks, robust ASR, deep learning 1.

We propose a novel approach to acoustic modeling based on recent advances in sparse representations. The key idea in sparse coding is to compute a compressed local representation of a signal via an over-complete basis or dictionary that is learned in an unsupervised way. In this study, we compute the local representation on speech spectrogram as the raw “signal ” and use it as the local sparse code to perform a standard phone classification task. A linear classifier is used that directly receives the coding space for making the classification decision. The simplicity of the linear classifier allows us to assess whether the sparse representations are sufficiently rich to serve as effective acoustic features for discriminating speech classes. Our experiments demonstrate competitive error rates when compared to other shallow approaches. An examination of the dictionary learned in sparse feature extraction demonstrates meaningful acoustic-phonetic properties that are captured by a collection of the dictionary entries. Index Terms: sparse coding, acoustic modeling, phone recognition, acoustic-phonetic properties

Linear Support Vector Machines (SVMs) have become very popular in vision as part of state-of-the-art object recognition and other classification tasks but require high dimensional feature spaces for good performance. Deep learning methods can find more compact representations but current methods employ multilayer perceptrons that require solving a difficult, non-convex optimization problem. We propose a deep non-linear classifier whose layers are SVMs and which incorporates random projection as its core stacking element. Our method learns layers of linear SVMs recursively transforming the original data manifold through a random projection of the weak prediction computed from each layer. Our method scales as linear SVMs, does not rely on any kernel computations or nonconvex optimization, and exhibits better generalization ability than kernel-based SVMs. This is especially true when the number of training samples is smaller than the dimensionality of data, a common scenario in many real-world applications. The use of random projections is key to our method, as we show in the experiments section, in which we observe a consistent improvement over previous --often more complicated-- methods on several vision and speech benchmarks.

The speaker diarization system developed at the International Computer Science Institute (ICSI) has played a prominent role in the speaker diarization community, and many researchers in the rich transcription community have adopted methods and techniques developed for the ICSI speaker diarization engine. Although there have been many related publications over the years, previous articles only presented changes and improvements rather than a description of the full system. Attempting to replicate the ICSI speaker diarization system as a complete entity would require an extensive literature review, and might ultimately fail due to component description version mismatches. This paper therefore presents the first full conceptual description of the ICSI speaker diarization system as presented to the National Institute of Standards Technology Rich Transcription 2009 (NIST RT-09) evaluation, which consists of online and offline subsystems, multi-stream and single-stream implementations, and audio and audio-visual approaches. Some of the components, such as the online system, have not been previously described. The paper also includes all necessary preprocessing steps, such as Wiener filtering, speech activity detection and beamforming.

Speaker diarization is the task of determining “who spoke when?” in an audio or video recording that contains an unknown amount of speech and also an unknown number of speakers. Initially, it was proposed as a research topic related to automatic speech recognition, where speaker diarization serves as an upstream processing step. Over recent years, however, speaker diarization has become an important key technology for many tasks, such as navigation, retrieval, or higher level inference on audio data. Accordingly, many important improvements in accuracy and robustness have been reported in journals and conferences in the area. The application domains, from broadcast news, to lectures and meetings, vary greatly and pose different problems, such as having access to multiple microphones and multimodal information or overlapping speech. The most recent review of existing technology dates back to 2006 and focuses on the broadcast news domain. In this paper, we review the current state-of-the-art, focusing on research developed since 2006 that relates predominantly to speaker diarization for conference meetings. Finally, we present an analysis of speaker diarization performance as reported through the NIST Rich Transcription evaluations on meeting data and identify important areas for future research.

In this paper, we extend the work done on integrating multilayer perceptron (MLP) networks with HMM systems via the Tandem approach. In particular, we explore whether the use of Deep Belief Networks (DBN) adds any substantial gain over MLPs on the Aurora2 speech recognition task under mismatched noise conditions. Our findings suggest that DBNs outperform single layer MLPs under the clean condition, but the gains diminish as the noise level is increased. Furthermore, using MFCCs in conjunction with the posteriors from DBNs outperforms merely using single DBNs in low to moderate noise conditions. MFCCs, however, do not help for the high noise settings.

International audience

This article presents a low-latency speaker diarization system (“who is speaking now?”) based on a hybrid approach that combines a traditional offline speaker diarization system (“who spoke when?”) with an online speaker identification system. The system fulfills all requirements of the diarization task, i.e. it does not need any a-priori information about the input, including no specific speaker models. After an initialization phase the approach allows a low-latency decision on the current speaker with an accuracy that is close to the underlying offline diarization system. The article describes the approach, evaluates the robustness of the system, and analyzes the latency/accuracy trade-off. Index Terms :S peaker Diarization, online, incremental, hybrid

In this paper, we propose a discriminative extension to agglomerative hierarchical clustering, a typical technique for speaker diarization, that fits seamlessly with most state-of-the art diarization algorithms. We propose to use maximum mutual information using bootstrapping i.e., initial predictions are used as input for retraining of models in an unsupervised fashion. This article describes this new approach, analyzes its behavior, and presents results on the official NIST Rich Transcription datasets. We show an absolute improvement of 4% DER with respect to the generative approach baseline. We also observe a strong correlation between the original error and the amount of improvement, that is, the better our predicted labels are, the more gain we obtain from discriminative training, which we interpret as a strong indication for the high potential of the extension.

In this article we define a multimedia content analysis problem, which we call multimodal location estimation: Given a video/image/audio file, the task is to determine where it was recorded. A single indication, such as a unique landmark, might already pinpoint a location precisely. In most cases, however, a combination of evidence from the visual and the acoustic domain will only narrow down the set of possible answers. Therefore, approaches to tackle this task should be inherently multimedia. While the task is hard, in fact sometimes unsolvable, training data can be leveraged from the Internet in large amounts. Moreover, even partially successful automatic estimation of location opens up new possibilities in video content matching, archiving, and organization. It could revolutionize law enforcement and computer-aided intelligence agency work, especially since both semi-automatic and fully automatic approaches would be possible. In this article, we describe our idea of growing multimodal location estimation as a research field in the multimedia community. Based on examples and scenarios, we propose a multimedia approach to leverage cues from the visual and the acoustic portions of a video as well as from given metadata. We also describe experiments to estimate the amount of available training data that could potentially be used as publicly available infrastructure for research in this field. Finally, we present an initial set of results based on acoustic and visual cues and discuss the massive challenges involved and some possible paths to solutions.

We present an indoor localization application leveraging the sensing capabilities of current state of the art smart phones. To the best of our knowledge, our application is the first one to be implemented in smart phones and integrating both offline and online phases of fingerprinting, delivering an accuracy of up to 1.5 meters. In particular, we have studied the possibilities offered by WiFi radio, cellular communications radio, accelerometer and magnetometer, already embedded in smart phones, with the intention to build a multimodal solution for localization. We have also implemented a new approach for the statistical processing of radio signal strengths, showing that it can outperform existing deterministic techniques.

The following article shows how a state-of-the-art speaker diarization system can be improved by combining traditional short-term features (MFCCs) with prosodic and other long-term features. First, we present a framework to study the speaker discriminability of 70 different long-term features. Then, we show how the top-ranked long-term features can be combined with short-term features to increase the accuracy of speaker diarization. The results were measured on standardized data sets (NIST RT) and show a consistent improvement of about 30% relative in diarization error rate compared to the best system presented at the NIST evaluation in 2007. This result was also verified on a wide set of meetings, which we call CombDev, that contains 21 meetings from previous evaluations. Since the prosodic and long-term features were selected using a diarization-independent speaker-discriminability study, we are confident that the same features are able to improve other systems that perform similar tasks

In this paper, we report our recent research aimed at improving the pronunciation-modeling component of a speech recognition system designed for mobile voice search. Our new discriminative learning technique overcomes the limitation of the traditional ways of introducing alternative pronunciations that often enlarge confusability across different lexical items. Instead, we make use of a phonetic recognizer to generate pronunciation candidates, which are then evaluated and selected using the global minimum-classification-error measure, guaranteeing a reduction of the training-set error rate after introducing alternative pronunciations. A maximum entropy approach is subsequently used to learn the weight parameters of the selected pronunciation candidates. Our experimental results demonstrate the effectiveness of the discriminative pronunciation learning technique in a real-world speech recognition task where pronunciation of business names presents special difficulty for high-accuracy speech recognition.

Speaker diarization is defined as the task of determining ldquowho spoke whenrdquo given an audio track and no other prior knowledge of any kind. The following article shows how a state-of-the-art speaker diarization system can be improved by combining traditional short-term features (MFCCs) with prosodic and other long-term features. First, we present a framework to study the speaker discriminability of 70 different long-term features. Then, we show how the top-ranked long-term features can be combined with short-term features to increase the accuracy of speaker diarization. The results were measured on standardized datasets (NIST RT) and show a consistent improvement of about 30% relative in diarization error rate compared to the best system presented at the NIST evaluation in 2007.

State-of-the-art speaker diarization systems for meetings are now at a point where overlapped speech contributes significantly to the errors made by the system. However, little if no work has yet been done on detecting overlapped speech. We present our initial work toward developing an overlap detection system for improved meeting diarization. We investigate various features, with a focus on high-precision performance for use in the detector, and examine performance results on a subset of the AMI Meeting Corpus. For the high-quality signal case of a single mixed-headset channel signal, we demonstrate a relative improvement of about 7.4% DER over the baseline diarization system, while for the more challenging case of the single far-field channel signal relative improvement is 3.6%. We also outline steps towards improvement and moving beyond this initial phase.

We present an update to our initial work [1] on overlapped speech detection for improving speaker diarization. Specifi-cally, we describe the addition of new features and feature warp-ing techniques that improve segmenter and, consequently, di-arization performance. We also demonstrate improved diariza-tion performance by additionally using overlap segment infor-mation in a new diarization pre-processing step which excludes overlap segments from speaker clustering. On a subset of the AMI Meeting Corpus we show that this overlap exclusion step nearly triples the relative improvement of diarization error rate as compared to overlap segment post-processing alone. Index Terms: speaker diarization, overlap detection 1.

We propose the use of modulation spectrogram features in speaker diarization. These features carry longer term characteristics of the acoustic signals than the widely used MFCCs, thus providing potential improvement by using both features in combination. Using the state-of-the-art ICSI speaker diarization system, an improvement of 20.77 % relative DER is obtained on the NIST Rich Transcription 2007 task with respect to the MFCC only system. Index Terms: modulation spectrogram, speaker diarization 1.

Many multimedia applications rely on the computation of logarithms, for example, when estimating log-likelihoods for Gaussian Mixture Models. Knowing of the demand to compute logarithms and other basic math functions rapidly, many hardware manufacturers provide libraries to perform calculations in hardware. Of course, these libraries are especially popular for the use in computer vision or audio analysis algorithms where a large amounts of data have to be processed. A downside of using specialized hardware though is that it increases the investment cost and the user is forced to use the same hardware, which is especially cumbersome when algorithms optimized for different specialized hardware are to be combined. This article presents the realization of a novel platform-independent, fast C-language implementation of the logarithm function. The idea behind the approach is to take advantage of the large amount of cache available in current processors. The logarithm implementation is compared to the current state of the art and we demonstrate the practical use of the algorithm in an actual speech analysis application.

The following article describes our technical demonstration of an online speaker identification system for conversations. A laptop with an internal microphone is centrally placed in the table of a meeting room. The system is able to identify the current speaker independent of spoken text or language with a latency of about 1.5 seconds and an accuracy of about 85% (as evaluated against the NIST RT benchmark). A Java GUI shows the image of the current speaker along with a timeline containing past speakers. Speakers are added to the system's database using a one-minute training procedure.

During the past few years, speaker diarization has achieved satisfying accuracy in terms of speaker Diarization Error Rate (DER). The most successful approaches, based on agglomerative clustering, however, exhibit an inherent computational complexity which makes real-time processing, especially in combination with further processing steps, almost impossible. In this article we present a framework to speed up agglomerative clustering speaker diarization. The basic idea is to adopt a computationally cheap method to reduce the hypothesis space of the more expensive and accurate model selection via Bayesian Information Criterion (BIC). Two strategies based on the pitch-correlogram and the unscented-transform based approximation of KL-divergence are used independently as a fast-match approach to select the most likely clusters to merge. We performed the experiments using the existing ICSI speaker diarization system. The new system using KL-divergence fast-match strategy only performs 14% of total BIC comparisons needed in the baseline system, speeds up the system by 41% without affecting the speaker Diarization Error Rate (DER). The result is a robust and faster than real-time speaker diarization system.

Kernel machines (e.g. SVM, KLDA) have shown state-of-the-art performance in several visual classification tasks. The classification performance of kernel machines greatly depends on the choice of kernels and its parameters. In this paper, we propose a method to search over a space of parameterized kernels using a gradient-descent based method. Our method effectively learns a non-linear representation of the data useful for classification and simultaneously performs dimensionality reduction. In addition, we suggest a new matrix formulation that simplifies and unifies previous approaches. The effectiveness and robustness of the proposed algorithm is demonstrated in both synthetic and real examples of pedestrian and mouth detection in images.