We introduce SMALLTALK LM, an innovative method for training a mixture of language models in an almost asynchronous manner. Each model of the mixture specializes in distinct parts of the data distribution, without the need of high-bandwidth communication between the nodes training each model. At inference, a lightweight router directs a given sequence to a single expert, according to a short prefix. This inference scheme naturally uses a fraction of the parameters from the overall mixture model. Unlike prior works on asynchronous LLM training, our routing method does not rely on full corpus clustering or access to metadata, making it more suitable for real-world applications. Our experiments on language modeling demonstrate that SMALLTALK LM achieves significantly lower perplexity than dense model baselines for the same total training FLOPs and an almost identical inference cost. Finally, in our downstream evaluations we outperform the dense baseline on 75% of the tasks.

Self-training (ST) and self-supervised learning (SSL) methods have demonstrated strong improvements in automatic speech recognition (ASR). In spite of these advances, to the best of our knowledge, there is no analysis of how the composition of the labelled and unlabelled datasets used in these methods affects the results. In this work we aim to analyse the effect of number of speakers in the training data on a recent SSL algorithm (wav2vec 2.0), and a recent ST algorithm (slimIPL). We perform a systematic analysis on both labeled and unlabeled data by varying the number of speakers while keeping the number of hours fixed and vice versa. Our findings suggest that SSL requires a large amount of unlabeled data to produce high accuracy results, while ST requires a sufficient number of speakers in the labelled data, especially in the low-regime setting. In this manner these two approaches improve supervised learning in different regimes of data composition.

Semi-supervised learning through pseudo-labeling has become a staple of state-of-the-art monolingual speech recognition systems. In this work, we extend pseudo-labeling to massively multilingual speech recognition with 60 languages. We propose a simple pseudo-labeling recipe that works well even with low-resource languages: train a supervised multilingual model, fine-tune it with semi-supervised learning on a target language, generate pseudo-labels for that language, and train a final model using pseudo-labels for all languages, either from scratch or by fine-tuning. Experiments on the labeled Common Voice and unlabeled VoxPopuli datasets show that our recipe can yield a model with better performance for many languages that also transfers well to LibriSpeech.

In this paper, we study training of automatic speech recognition system in a weakly supervised setting where the order of words in transcript labels of the audio training data is not known. We train a word-level acoustic model which aggregates the distribution of all output frames using LogSumExp operation and uses a cross-entropy loss to match with the ground-truth words distribution. Using the pseudo-labels generated from this model on the training set, we then train a letter-based acoustic model using Connectionist Temporal Classification loss. Our system achieves 2.3%/4.6% on test-clean/test-other subsets of LibriSpeech, which closely matches with the supervised baseline's performance.

As the computational requirements for machine learning systems and the size and complexity of machine learning frameworks increases, essential framework innovation has become challenging. While computational needs have driven recent compiler, networking, and hardware advancements, utilization of those advancements by machine learning tools is occurring at a slower pace. This is in part due to the difficulties involved in prototyping new computational paradigms with existing frameworks. Large frameworks prioritize machine learning researchers and practitioners as end users and pay comparatively little attention to systems researchers who can push frameworks forward — we argue that both are equally important stakeholders. We introduce Flashlight, an open-source library built to spur innovation in machine learning tools and systems by prioritizing open, modular, customizable internals and state-of-the-art, research-ready models and training setups across a variety of domains. Flashlight allows systems researchers to rapidly prototype and experiment with novel ideas in machine learning computation and has low overhead, competing with and often outperforming other popular machine learning frameworks. We see Flashlight as a tool enabling research that can benefit widely used libraries downstream and bring machine learning and systems researchers closer together.

We develop an algorithm which can learn from partially labeled and unsegmented sequential data. Most sequential loss functions, such as Connectionist Temporal Classification (CTC), break down when many labels are missing. We address this problem with Star Temporal Classification (STC) which uses a special star token to allow alignments which include all possible tokens whenever a token could be missing. We express STC as the composition of weighted finite-state transducers (WFSTs) and use GTN (a framework for automatic differentiation with WFSTs) to compute gradients. We perform extensive experiments on automatic speech recognition. These experiments show that STC can close the performance gap with supervised baseline to about 1% WER when up to 70% of the labels are missing. We also perform experiments in handwriting recognition to show that our method easily applies to other temporal classification tasks.

In this paper, we introduce the Kaizen framework that uses a continuously improving teacher to generate pseudo-labels for semi-supervised speech recognition (ASR). The proposed approach uses a teacher model which is updated as the exponential moving average (EMA) of the student model parameters. We demonstrate that it is critical for EMA to be accumulated with full-precision floating point. The Kaizen framework can be seen as a continuous version of the iterative pseudo-labeling approach for semi-supervised training. It is applicable for different training criteria, and in this paper we demonstrate its effectiveness for frame-level hybrid hidden Markov model-deep neural network (HMM-DNN) systems as well as sequence-level Connectionist Temporal Classification (CTC) based models. For large scale real-world unsupervised public videos in UK English and Italian languages the proposed approach i) shows more than 10% relative word error rate (WER) reduction over standard teacher-student training; ii) using just 10 hours of supervised data and a large amount of unsupervised data closes the gap to the upper-bound supervised ASR system that uses 650h or 2700h respectively.

Self-supervised learning (SSL) has shown promise in learning representations of audio that are useful for automatic speech recognition (ASR). But, training SSL models like wav2vec 2.0 requires a two-stage pipeline. In this paper we demonstrate a single-stage training of ASR models that can utilize both unlabeled and labeled data. During training, we alternately minimize two losses: an unsupervised masked Contrastive Predictive Coding (CPC) loss and the supervised audio-to-text alignment loss Connectionist Temporal Classification (CTC). We show that this joint training method directly optimizes performance for the downstream ASR task using unsupervised data while achieving similar word error rates to wav2vec 2.0 on the Librispeech 100-hours dataset. Finally, we postulate that solving the contrastive task is a regularization for the supervised CTC loss.

Self-training and unsupervised pre-training have emerged as effective approaches to improve speech recognition systems using unlabeled data. However, it is not clear whether they learn similar patterns or if they can be effectively combined. In this paper, we show that pseudo-labeling and pre-training with wav2vec 2.0 are complementary in a variety of labeled data setups. Using just 10 minutes of labeled data from Libri-light as well as 53k hours of unlabeled data from LibriVox achieves word error rates (WER) of 2.8%/4.8% on the clean and other test sets of Librispeech – rivaling the best published systems trained on 960 hours of labeled data only a year ago. Training on all labeled data of Librispeech achieves WERs of 1.5%/3.1%.

This paper presents XLSR which learns cross-lingual speech representations by pretraining a single model from the raw waveform of speech in multiple languages. We build on wav2vec 2.0 which is trained by solving a contrastive task over masked latent speech representations and jointly learns a quantization of the latents shared across languages. The resulting model is fine-tuned on labeled data and experiments show that cross-lingual pretraining significantly outperforms monolingual pretraining. On the CommonVoice benchmark, XLSR shows a relative phoneme error rate reduction of 72% compared to the best known results. On BABEL, our approach improves word error rate by 16% relative compared to a comparable system. Our approach enables a single multilingual speech recognition model which is competitive to strong individual models. Analysis shows that the latent discrete speech representations are shared across languages with increased sharing for related languages. We hope to catalyze research in low-resource speech understanding by releasing XLSR-53, a large model pretrained in 53 languages.

Self-supervised learning of speech representations has been a very active research area but most work is focused on a single domain such as read audio books for which there exist large quantities of labeled and unlabeled data. In this paper, we explore more general setups where the domain of the unlabeled data for pre-training data differs from the domain of the labeled data for fine-tuning, which in turn may differ from the test data domain. Our experiments show that using target domain data during pre-training leads to large performance improvements across a variety of setups. On a large-scale competitive setup, we show that pre-training on unlabeled in-domain data reduces the gap between models trained on in-domain and out-of-domain labeled data by 66%-73%. This has obvious practical implications since it is much easier to obtain unlabeled target domain data than labeled data. Moreover, we find that pre-training on multiple domains improves generalization performance on domains not seen during training. Code and models will be made available at https://github.com/pytorch/fairseq.

Recent results in end-to-end automatic speech recognition have demonstrated the efficacy of pseudo-labeling for semi-supervised models trained both with Connectionist Temporal Classification (CTC) and Sequence-to-Sequence (seq2seq) losses. Iterative Pseudo-Labeling (IPL), which continuously trains a single model using pseudo-labels iteratively re-generated as the model learns, has been shown to further improve performance in ASR. We improve upon the IPL algorithm: as the model learns, we propose to iteratively re-generate transcriptions with hard labels (the most probable tokens), that is, without a language model. We call this approach Language-Model-Free IPL (slimIPL) and give a resultant training setup for low-resource settings with CTC-based models. slimIPL features a dynamic cache for pseudo-labels which reduces sensitivity to changes in relabeling hyperparameters and results in improves training stability. slimIPL is also highly-efficient and requires 3.5-4x fewer computational resources to converge than other state-of-the-art semi/self-supervised approaches. With only 10 hours of labeled audio, slimIPL is competitive with self-supervised approaches, and is state-of-the-art with 100 hours of labeled audio without the use of a language model both at test time and during pseudo-label generation.

Is pushing numbers on a single benchmark valuable in automatic speech recognition? Research results in acoustic modeling are typically evaluated based on performance on a single dataset. While the research community has coalesced around various benchmarks, we set out to understand generalization performance in acoustic modeling across datasets - in particular, if models trained on a single dataset transfer to other (possibly out-of-domain) datasets. We show that, in general, reverberative and additive noise augmentation improves generalization performance across domains. Further, we demonstrate that when a large enough set of benchmarks is used, average word error rate (WER) performance over them provides a good proxy for performance on real-world noisy data. Finally, we show that training a single acoustic model on the most widely-used datasets - combined - reaches competitive performance on both research and real-world benchmarks.

Without positional information, attention-based Transformer neural networks are permutation-invariant. Absolute or relative positional embeddings are the most popular ways to feed Transformer models with positional information. Absolute positional embeddings are simple to implement, but suffer from generalization issues when evaluating on sequences longer than seen at training time. Relative positions are more robust to input length change, but are more complex to implement and yield inferior model throughput due to extra computational and memory costs. In this paper, we propose an augmentation-based approach (CAPE) for absolute positional embeddings, which keeps the advantages of both absolute (simplicity and speed) and relative positional embeddings (better generalization). In addition, our empirical evaluation on state-of-the-art models in machine translation, image and speech recognition demonstrates that CAPE leads to better generalization performance as well as increased stability with respect to training hyper-parameters.

We introduce a new collection of spoken English audio suitable for training speech recognition systems under limited or no supervision. It is derived from open-source audio books from the LibriVox project. It contains over 60K hours of audio, which is, to our knowledge, the largest freely-available corpus of speech. The audio has been segmented using voice activity detection and is tagged with SNR, speaker ID and genre descriptions. Additionally, we provide baseline systems and evaluation metrics working under three settings: (1) the zero resource/unsupervised setting (ABX), (2) the semi- supervised setting (PER, CER) and (3) the distant supervision setting (WER). Settings (2) and (3) use limited textual resources (10 minutes to 10 hours) aligned with the speech. Setting (3) uses large amounts of unaligned text. They are evaluated on the standard LibriSpeech dev and test sets for comparison with the supervised state-of-the-art.

We propose a direct-to-word sequence model which uses a word network to learn word embeddings from letters. The word network can be integrated seamlessly with arbitrary sequence models including Connectionist Temporal Classification and encoder-decoder models with attention. We show our direct-to-word model can achieve word error rate gains over sub-word level models for speech recognition. We also show that our direct-to-word approach retains the ability to predict words not seen at training time without any retraining. Finally, we demonstrate that a word-level model can use a larger stride than a sub-word level model while maintaining accuracy. This makes the model more efficient both for training and inference.

We study training a single acoustic model for multiple languages with the aim of improving automatic speech recognition (ASR) performance on low-resource languages, and over-all simplifying deployment of ASR systems that support diverse languages. We perform an extensive benchmark on 51 languages, with varying amount of training data by language(from 100 hours to 1100 hours). We compare three variants of multilingual training from a single joint model without knowing the input language, to using this information, to multiple heads (one per language cluster). We show that multilingual training of ASR models on several languages can improve recognition performance, in particular, on low resource languages. We see 20.9%, 23% and 28.8% average WER relative reduction compared to monolingual baselines on joint model, joint model with language input and multi head model respectively. To our knowledge, this is the first work studying multilingual ASR at massive scale, with more than 50 languages and more than 16,000 hours of audio across them.

We design an online end-to-end speech recognition system based on Time-Depth Separable (TDS) convolutions and Connectionist Temporal Classification (CTC). We improve the core TDS architecture in order to limit the future context and hence reduce latency while maintaining accuracy. The system has almost three times the throughput of a well tuned hybrid ASR baseline while also having lower latency and a better word error rate. Also important to the efficiency of the recognizer is our highly optimized beam search decoder. To show the impact of our design choices, we analyze throughput, latency, accuracy, and discuss how these metrics can be tuned based on the user requirements.

This paper introduces Multilingual LibriSpeech (MLS) dataset, a large multilingual corpus suitable for speech research. The dataset is derived from read audiobooks from LibriVox and consists of 8 languages, including about 44.5K hours of English and a total of about 6K hours for other languages. Additionally, we provide Language Models (LM) and baseline Automatic Speech Recognition (ASR) models and for all the languages in our dataset. We believe such a large transcribed dataset will open new avenues in ASR and Text-To-Speech (TTS) research. The dataset will be made freely available for anyone at http://www.openslr.org.

Pseudo-labeling has recently shown promise in end-to-end automatic speech recognition (ASR).We study Iterative Pseudo-Labeling (IPL), a semi-supervised algorithm which efficiently performs multiple iterations of pseudo-labeling on unlabeled data as the acoustic model evolves.In particular, IPL fine tunes an existing model at each iteration using both labeled data and a subset of unlabeled data.We study the main components of IPL: decoding with a language model and data augmentation.We then demonstrate the effectiveness of IPL by achieving state-of-the-art word-error rate on the LIBRISPEECH test sets in both standard and low-resource setting.We also study the effect of language models trained on different corpora to show IPL can effectively utilize additional text.Finally, we release a new large in-domain text corpus which does not overlap with the LIBRISPEECH training transcriptions to foster research in low-resource, semi-supervised ASR.

Transcribed datasets typically contain speaker identity for each instance in the data. We investigate two ways to incorporate this information during training: Multi-Task Learning and Adversarial Learning. In multi-task learning, the goal is speaker prediction; we expect a performance improvement with this joint training if the two tasks of speech recognition and speaker recognition share a common set of underlying features. In contrast, adversarial learning is a means to learn representations invariant to the speaker. We then expect better performance if this learnt invariance helps generalizing to new speakers. While the two approaches seem natural in the context of speech recognition, they are incompatible because they correspond to opposite gradients back-propagated to the model. In order to better understand the effect of these approaches in terms of error rates, we compare both strategies in controlled settings. Moreover, we explore the use of additional un-transcribed data in a semi-supervised, adversarial learning manner to improve error rates. Our results show that deep models trained on big datasets already develop invariant representations to speakers without any auxiliary loss. When considering adversarial learning and multi-task learning, the impact on the acoustic model seems minor. However, models trained in a semi-supervised manner can improve error-rates.

This paper introduces wav2letter++, a fast open-source deep learning speech recognition framework. wav2letter++ is written entirely in C++, and uses the ArrayFire tensor library for maximum efficiency. We explain the architecture and design of the wav2letter++ system and compare it to other major open-source speech recognition systems. In some cases wav2letter++ is more than 2× faster than other optimized frameworks for training end-to-end neural networks for speech recognition. We also show that wav2letter++ training times scale linearly to 64 GPUs, the most we tested, for models with 100 million parameters. High-performance frameworks enable fast iteration, which is often a crucial factor in successful research and model tuning on new datasets and tasks.

We introduce a new beam search decoder that is fully differentiable, making it possible to optimize at training time through the inference procedure. Our decoder allows us to combine models which operate at different granularities (e.g. acoustic and language models). It can be used when target sequences are not aligned to input sequences by considering all possible alignments between the two. We demonstrate our approach scales by applying it to speech recognition, jointly training acoustic and word-level language models. The system is end-to-end, with gradients flowing through the whole architecture from the word-level transcriptions. Recent research efforts have shown that deep neural networks with attention-based mechanisms can successfully train an acoustic model from the final transcription, while implicitly learning a language model. Instead, we show that it is possible to discriminatively train an acoustic model jointly with an explicit and possibly pre-trained language model.

We propose a fully convolutional sequence-to-sequence encoder architecture with a simple and efficient decoder. Our model improves WER on LibriSpeech while being an order of magnitude more efficient than a strong RNN baseline. Key to our approach is a time-depth separable convolution block which dramatically reduces the number of parameters in the model while keeping the receptive field large. We also give a stable and efficient beam search inference procedure which allows us to effectively integrate a language model. Coupled with a convolutional language model, our time-depth separable convolution architecture improves by more than 22% relative WER over the best previously reported sequence-to-sequence results on the noisy LibriSpeech test set.

Lexicon-free speech recognition naturally deals with the problem of out-of-vocabulary (OOV) words. In this paper, we show that character-based language models (LM) can perform as well as word-based LMs for speech recognition, in word error rates (WER), even without restricting the decoding to a lexicon. We study character-based LMs and show that convolutional LMs can effectively leverage large (character) contexts, which is key for good speech recognition performance downstream. We specifically show that the lexicon-free decoding performance (WER) on utterances with OOV words using character-based LMs is better than lexicon-based decoding, both with character or word-based LMs.

We explore unsupervised pre-training for speech recognition by learning representations of raw audio. wav2vec is trained on large amounts of unlabeled audio data and the resulting representations are then used to improve acoustic model training. We pre-train a simple multi-layer convolutional neural network optimized via a noise contrastive binary classification task. Our experiments on WSJ reduce WER of a strong character-based log-mel filterbank baseline by up to 36% when only a few hours of transcribed data is available. Our approach achieves 2.43% WER on the nov92 test set. This outperforms Deep Speech 2, the best reported character-based system in the literature while using two orders of magnitude less labeled training data.

State-of-the-art speech recognition systems rely on fixed, hand-crafted features such as mel-filterbanks to preprocess the waveform before the training pipeline. In this paper, we study end-to-end systems trained directly from the raw waveform, building on two alternatives for trainable replacements of mel-filterbanks that use a convolutional architecture. The first one is inspired by gammatone filterbanks (Hoshen et al., 2015; Sainath et al, 2015), and the second one by the scattering transform (Zeghidour et al., 2017). We propose two modifications to these architectures and systematically compare them to mel-filterbanks, on the Wall Street Journal dataset. The first modification is the addition of an instance normalization layer, which greatly improves on the gammatone-based trainable filterbanks and speeds up the training of the scattering-based filterbanks. The second one relates to the low-pass filter used in these approaches. These modifications consistently improve performances for both approaches, and remove the need for a careful initialization in scattering-based trainable filterbanks. In particular, we show a consistent improvement in word error rate of the trainable filterbanks relatively to comparable mel-filterbanks. It is the first time end-to-end models trained from the raw signal significantly outperform mel-filterbanks on a large vocabulary task under clean recording conditions.

This paper aims to classify and locate objects accurately and efficiently, without using bounding box annotations. It is challenging as objects in the wild could appear at arbitrary locations and in different scales. In this paper, we propose a novel classification architecture ProNet based on convolutional neural networks. It uses computationally efficient neural networks to propose image regions that are likely to contain objects, and applies more powerful but slower networks on the proposed regions. The basic building block is a multi-scale fully-convolutional network which assigns object confidence scores to boxes at different locations and scales. We show that such networks can be trained effectively using image-level annotations, and can be connected into cascades or trees for efficient object classification. ProNet outperforms previous state-of-the-art significantly on PASCAL VOC 2012 and MS COCO datasets for object classification and point-based localization.

We are interested in inferring object segmentation by leveraging only object class information, and by considering only minimal priors on the object segmentation task. This problem could be viewed as a kind of weakly supervised segmentation task, and naturally fits the Multiple Instance Learning (MIL) framework: every training image is known to have (or not) at least one pixel corresponding to the image class label, and the segmentation task can be rewritten as inferring the pixels belonging to the class of the object (given one image, and its object class). We propose a Convolutional Neural Network-based model, which is constrained during training to put more weight on pixels which are important for classifying the image. We show that at test time, the model has learned to discriminate the right pixels well enough, such that it performs very well on an existing segmentation benchmark, by adding only few smoothing priors. Our system is trained using a subset of the Imagenet dataset and the segmentation experiments are performed on the challenging Pascal VOC dataset (with no fine-tuning of the model on Pascal VOC). Our model beats the state of the art results in weakly supervised object segmentation task by a large margin. We also compare the performance of our model with state of the art fully-supervised segmentation approaches.

State-of-the-art automatic speech recognition systems model the relationship between acoustic speech signal and phone classes in two stages, namely, extraction of spectral-based features based on prior knowledge followed by training of acoustic model, typically an artificial neural network (ANN). In our recent work, it was shown that Convolutional Neural Networks (CNNs) can model phone classes from raw acoustic speech signal, reaching performance on par with other existing feature-based approaches. This paper extends the CNN-based approach to large vocabulary speech recognition task. More precisely, we compare the CNN-based approach against the conventional ANN-based approach on Wall Street Journal corpus. Our studies show that the CNN-based approach achieves better performance than the conventional ANN-based approach with as many parameters. We also show that the features learned from raw speech by the CNN-based approach could generalize across different databases.

Generating a novel textual description of an image is an interesting problem that connects computer vision and natural language processing. In this paper, we present a simple model that is able to generate descriptive sentences given a sample image. This model has a strong focus on the syntax of the descriptions. We train a purely linear model to embed an image representation (generated from a previously trained Convolutional Neural Network) into a multimodal space that is common to the images and the phrases that are used to described them. The system is then able to infer phrases from a given image sample. Based on the sentence description statistics, we propose a simple language model that can produce relevant descriptions for a given test image using the phrases inferred. Our approach, which is considerably simpler than state-of-the-art models, achieves comparable results in two popular datasets for the task: Flickr30k and the recently proposed Microsoft COCO.

Automatic speech recognition systems typically model the relationship between the acoustic speech signal and the phones in two separate steps: feature extraction and classifier training.In our recent works, we have shown that, in the framework of convolutional neural networks (CNN), the relationship between the raw speech signal and the phones can be directly modeled and ASR systems competitive to standard approach can be built.In this paper, we first analyze and show that, between the first two convolutional layers, the CNN learns (in parts) and models the phone-specific spectral envelope information of 2-4 ms speech.Given that we show that the CNN-based approach yields ASR trends similar to standard short-term spectral based ASR system under mismatched (noisy) conditions, with the CNN-based approach being more robust.

Recent object detection systems rely on two critical steps: (1) a set of object proposals is predicted as efficiently as possible, and (2) this set of candidate proposals is then passed to an object classifier. Such approaches have been shown they can be fast, while achieving the state of the art in detection performance. In this paper, we propose a new way to generate object proposals, introducing an approach based on a discriminative convolutional network. Our model is trained jointly with two objectives: given an image patch, the first part of the system outputs a class-agnostic segmentation mask, while the second part of the system outputs the likelihood of the patch being centered on a full object. At test time, the model is efficiently applied on the whole test image and generates a set of segmentation masks, each of them being assigned with a corresponding object likelihood score. We show that our model yields significant improvements over state-of-the-art object proposal algorithms. In particular, compared to previous approaches, our model obtains substantially higher object recall using fewer proposals. We also show that our model is able to generalize to unseen categories it has not seen during training. Unlike all previous approaches for generating object masks, we do not rely on edges, superpixels, or any other form of low-level segmentation.

Word embeddings resulting from neural language models have been shown to be successful for a large variety of NLP tasks. However, such architecture might be difficult to train and time-consuming. Instead, we propose to drastically simplify the word embeddings computation through a Hellinger PCA of the word co-occurence matrix. We compare those new word embeddings with the Collobert and Weston (2008) embeddings on several NLP tasks and show that we can reach similar or even better performance. 1

The goal of the scene labeling task is to assign a class label to each pixel in an image. To ensure a good visual coherence and a high class accuracy, it is essential for a model to capture long range pixel) label dependencies in images. In a feed-forward architecture, this can be achieved simply by considering a sufficiently large input context patch, around each pixel to be labeled. We propose an approach that consists of a recurrent convolutional neural network which allows us to consider a large input context while limiting the capacity of the model. Contrary to most standard approaches, our method does not rely on any segmentation technique nor any task-specific features. The system is trained in an end-to-end manner over raw pixels, and models complex spatial dependencies with low inference cost. As the context size increases with the built-in recurrence, the system identifies and corrects its own errors. Our approach yields state-of-the-art performance on both the Stanford Background Dataset and the SIFT Flow Dataset, while remaining very fast at test time.

In hybrid hidden Markov model/artificial neural networks (HMM/ANN) automatic speech recognition (ASR) system, the phoneme class conditional probabilities are estimated by first extracting acoustic features from the speech signal based on prior knowledge such as, speech perception or/and speech production knowledge, and, then modeling the acoustic features with an ANN. Recent advances in machine learning techniques, more specifically in the field of image processing and text processing, have shown that such divide and conquer strategy (i.e., separating feature extraction and modeling steps) may not be necessary. Motivated from these studies, in the framework of convolutional neural networks (CNNs), this paper investigates a novel approach, where the input to the ANN is raw speech signal and the output is phoneme class conditional probability estimates. On TIMIT phoneme recognition task, we study different ANN architectures to show the benefit of CNNs and compare the proposed approach against conventional approach where, spectral-based feature MFCC is extracted and modeled by a multilayer perceptron. Our studies show that the proposed approach can yield comparable or better phoneme recognition performance when compared to the conventional approach. It indicates that CNNs can learn features relevant for phoneme classification automatically from the raw speech signal.

Many Knowledge Bases (KBs) are now readily available and encompass colossal quantities of information thanks to either a long-term funding effort (e.g. WordNet, OpenCyc) or a collaborative process (e.g. Freebase, DBpedia). However, each of them is based on a different rigorous symbolic framework which makes it hard to use their data in other systems. It is unfortunate because such rich structured knowledge might lead to a huge leap forward in many other areas of AI like nat- ural language processing (word-sense disambiguation, natural language understanding, ...), vision (scene classification, image semantic annotation, ...) or collaborative filtering. In this paper, we present a learning process based on an innovative neural network architecture designed to embed any of these symbolic representations into a more flexible continuous vector space in which the original knowledge is kept and enhanced. These learnt embeddings would allow data from any KB to be easily used in recent machine learning meth- ods for prediction and information retrieval. We illustrate our method on WordNet and Freebase and also present a way to adapt it to knowledge extraction from raw text.

In this article we propose Supervised Semantic Indexing (SSI), an algorithm that is trained on (query, document) pairs of text documents to predict the quality of their match. Like Latent Semantic Indexing (LSI), our models take account of correlations between words (synonymy, polysemy). However, unlike LSI our models are trained with a supervised signal directly on the ranking task of interest, which we argue is the reason for our superior results. As the query and target texts are modeled separately, our approach is easily generalized to different retrieval tasks, such as online advertising placement. Dealing with models on all pairs of words features is computationally challenging. We propose several improvements to our basic model for addressing this issue, including low rank (but diagonal preserving) representations, and correlated feature hashing (CFH). We provide an empirical study of all these methods on retrieval tasks based on Wikipedia documents as well as an Internet advertisement task. We obtain state-of-the-art performance while providing realistically scalable methods.

We describe a novel simple and highly scalable semi-supervised method called Word-Class Distribution Learning (WCDL), and apply it task of information extraction (IE) by utilizing unlabeled sentences to improve supervised classification methods. WCDL iteratively builds class label distributions for each word in the dictionary by averaging predicted labels over all cases in the unlabeled corpus, and re-training a base classifier adding these distributions as word features. In contrast, traditional self-training or co-training methods self-labeled examples (rather than features) which can degrade performance due to incestuous learning bias. WCDL exhibits robust behavior, and has no difficult parameters to tune. We applied our method on German and English name entity recognition (NER) tasks. WCDL shows improvements over self-training, multi-task semi-supervision or supervision alone, in particular yielding a state-of-the art 75.72 F1 score on the German NER task.

Typical information extraction (IE) systems can be seen as tasks assigning labels to words in a natural language sequence. The performance is restricted by the availability of labeled words. To tackle this issue, we propose a semi-supervised approach to improve the sequence labeling procedure in IE through a class of algorithms with self-learned features (SLF). A supervised classifier can be trained with annotated text sequences and used to classify each word in a large set of unannotated sentences. By averaging predicted labels over all cases in the unlabeled corpus, SLF training builds class label distribution patterns for each word (or word attribute) in the dictionary and re-trains the current model iteratively adding these distributions as extra word features. Basic SLF models how likely a word could be assigned to target class types. Several extensions are proposed, such as learning words' class boundary distributions. SLF exhibits robust and scalable behaviour and is easy to tune. We applied this approach on four classical IE tasks: named entity recognition (German and English), part-of-speech tagging (English) and one gene name recognition corpus. Experimental results show effective improvements over the supervised baselines on all tasks. In addition, when compared with the closely related self-training idea, this approach shows favorable advantages.

Humans and animals learn much better when the examples are not randomly presented but organized in a meaningful order which illustrates gradually more concepts, and gradually more complex ones. Here, we formalize such training strategies in the context of machine learning, and call them "curriculum learning". In the context of recent research studying the difficulty of training in the presence of non-convex training criteria (for deep deterministic and stochastic neural networks), we explore curriculum learning in various set-ups. The experiments show that significant improvements in generalization can be achieved. We hypothesize that curriculum learning has both an effect on the speed of convergence of the training process to a minimum and, in the case of non-convex criteria, on the quality of the local minima obtained: curriculum learning can be seen as a particular form of continuation method (a general strategy for global optimization of non-convex functions).

This work proposes a learning method for deep architectures that takes advantage of sequential data, in particular from the temporal coherence that naturally exists in unlabeled video recordings. That is, two successive frames are likely to contain the same object or objects. This coherence is used as a supervisory signal over the unlabeled data, and is used to improve the performance on a supervised task of interest. We demonstrate the effectiveness of this method on some pose invariant object and face recognition tasks. 1.

We present a class of nonlinear (polynomial) models that are discriminatively trained to directly map from the word content in a query-document or document-document pair to a ranking score. Dealing with polynomial models on word features is computationally challenging. We propose a low rank (but diagonal preserving) representation of our polynomial models to induce feasible memory and computation requirements. We provide an empirical study on retrieval tasks based on Wikipedia documents, where we obtain state-of-the-art performance while providing realistically scalable methods.

We describe a single convolutional neural network architecture that, given a sentence, outputs a host of language processing predictions: part-of-speech tags, chunks, named entity tags, semantic roles, semantically similar words and the likelihood that the sentence makes sense (grammatically and semantically) using a language model. The entire network is trained jointly on all these tasks using weight-sharing, an instance of multitask learning. All the tasks use labeled data except the language model which is learnt from unlabeled text and represents a novel form of semi-supervised learning for the shared tasks. We show how both multitask learning and semi-supervised learning improve the generalization of the shared tasks, resulting in state-of-the-art-performance.

We show how the regularizer of Transductive Support Vector Machines (TSVM) can be trained by stochastic gradient descent for linear models and multi-layer architectures. The resulting methods can be trained online, have vastly superior training and testing speed to existing TSVM algorithms, can encode prior knowledge in the network architecture, and obtain competitive error rates. We then go on to propose a natural generalization of the TSVM loss function that takes into account neighborhood and manifold information directly, unifying the two-stage Low Density Separation method into a single criterion, and leading to state-of-the-art results.

Abstract. We show how nonlinear embedding algorithms popular for use with “shallow ” semi-supervised learning techniques such as kernel methods can be easily applied to deep multi-layer architectures, either as a regularizer at the output layer, or on each layer of the architecture. This trick provides a simple alternative to existing approaches to deep learning whilst yielding competitive error rates compared to those methods, and existing shallow semi-supervised techniques.

Convex learning algorithms, such as Support Vector Machines (SVMs), are often seen as highly desirable because they offer strong practical properties and are amenable to theoretical analysis. However, in this work we show how non-convexity can provide scalability advantages over convexity. We show how concave-convex programming can be applied to produce (i) faster SVMs where training errors are no longer support vectors, and (ii) much faster Transductive SVMs.

In this paper we study a new framework introduced by Vapnik (1998) and Vapnik (2006) that is an alternative capacity concept to the large margin approach. In the particular case of binary classification, we are given a set of labeled examples, and a collection of "non-examples" that do not belong to either class of interest. This collection, called the Universum, allows one to encode prior knowledge by representing meaningful concepts in the same domain as the problem at hand. We describe an algorithm to leverage the Universum by maximizing the number of observed contradictions, and show experimentally that this approach delivers accuracy improvements over using labeled data alone.

We show how the concave-convex procedure can be applied to transductive SVMs, which traditionally require solving a combinatorial search problem. This provides for the first time a highly scalable algorithm in the nonlinear case. Detailed experiments verify the utility of our approach. Software is available at http://www.kyb.tuebingen.mpg.de/bs/people/fabee/transduction.html.

Several second-order optimization methods for gradient descent algorithms have been proposed over the years, but they usually need to compute the inverse of the Hessian of the cost function (or an approximation of this inverse) during training. In most cases, this leads to an O(n/sup 2/) cost in time and space per iteration, where n is the number of parameters, which is prohibitive for large n. We propose instead a study of the Hessian before training. Based on a second order analysis, we show that a block-diagonal Hessian yields an easier optimization problem than a full Hessian. We also show that the condition of block-diagonality in common machine learning models can be achieved by simply selecting an appropriate training criterion. Finally, we propose a version of the SVM criterion applied to MLPs, which verifies the aspects highlighted in this second order analysis, but also yields very good generalization performance in practice, taking advantage of the margin effect. Several empirical comparisons on two benchmark datasets are given to illustrate this approach.

We propose to study links between three important classification algorithms: Perceptrons, Multi-Layer Perceptrons (MLPs) and Support Vector Machines (SVMs). We first study ways to control the capacity of Perceptrons (mainly regularization parameters and early stopping), using the margin idea introduced with SVMs. After showing that under simple conditions a Perceptron is equivalent to an SVM, we show it can be computationally expensive in time to train an SVM (and thus a Perceptron) with stochastic gradient descent, mainly because of the margin maximization term in the cost function. We then show that if we remove this margin maximization term, the learning rate or the use of early stopping can still control the margin. These ideas are extended afterward to the case of MLPs. Moreover, under some assumptions it also appears that MLPs are a kind of mixture of SVMs, maximizing the margin in the hidden layer space. Finally, we present a very simple MLP based on the previous findings, which yields better performances in generalization and speed than the other models.

We present an overview of research at IDIAP on speech & face based biometric authentication. This paper covers user-customised passwords, adaptation techniques, confidence measures (for use in fusion of audio & visual scores), face verification in difficult image conditions, as well as other related research issues. We also overviewed the open source Torch library, which has aided in the implementation of the above mentioned techniques.

A challenge for statistical learning is to deal with large data sets, e.g. in data mining. The training time of ordinary Support Vector Machines is at least quadratic, which raises a serious research challenge if we want to deal with data sets of millions of examples. We propose a "hard parallelizable mixture" methodology which yields significantly reduced training time through modularization and parallelization: the training data is iteratively partitioned by a "gater" model in such a way that it becomes easy to learn an "expert" model separately in each region of the partition. A probabilistic extension and the use of a set of generative models allows representing the gater so that all pieces of the model are locally trained. For SVMs, time complexity appears empirically to local growth linearly with the number of examples, while generalization performance can be enhanced. For the probabilistic version of the algorithm, the iterative algorithm probably goes down in a cost function that is an upper bound on the negative log-likelihood.

Support vector machines (SVMs) are the state-of-the-art models for many classification problems, but they suffer from the complexity of their training algorithm, which is at least quadratic with respect to the number of examples. Hence, it is hopeless to try to solve real-life problems having more than a few hundred thousand examples with SVMs. This article proposes a new mixture of SVMs that can be easily implemented in parallel and where each SVM is trained on a small subset of the whole data set. Experiments on a large benchmark data set (Forest) yielded significant time improvement (time complexity appears empirically to locally grow linearly with the number of examples). In addition, and surprisingly, a significant improvement in generalization was observed.

Support Vector Machines (SVMs) are currently the state-of-the-art models for many classification problems but they suffer from the complexity of their train(cid:173) ing algorithm which is at least quadratic with respect to the number of examples. Hence, it is hopeless to try to solve real-life problems having more than a few hundreds of thousands examples with SVMs. The present paper proposes a new mixture of SVMs that can be easily implemented in parallel and where each SVM is trained on a small subset of the whole dataset. Experiments on a large benchmark dataset (Forest) as well as a difficult speech database, yielded significant time improvement (time complexity appears empirically to locally grow linearly with the number of examples) . In addition, and that is a surprise, a significant improvement in generalization was observed on Forest.