We introduce Hibiki, a decoder-only model for simultaneous speech translation. Hibiki leverages a multistream language model to synchronously process source and target speech, and jointly produces text and audio tokens to perform speech-to-text and speech-to-speech translation. We furthermore address the fundamental challenge of simultaneous interpretation, which unlike its consecutive counterpart --where one waits for the end of the source utterance to start translating-- adapts its flow to accumulate just enough context to produce a correct translation in real-time, chunk by chunk. To do so, we introduce a weakly-supervised method that leverages the perplexity of an off-the-shelf text translation system to identify optimal delays on a per-word basis and create aligned synthetic data. After supervised training, Hibiki performs adaptive, simultaneous speech translation with vanilla temperature sampling. On a French-English simultaneous speech translation task, Hibiki demonstrates state-of-the-art performance in translation quality, speaker fidelity and naturalness. Moreover, the simplicity of its inference process makes it compatible with batched translation and even real-time on-device deployment. We provide examples on *huggingface.co/spaces/kyutai/hibiki-samples* as well as models and inference code at *github.com/kyutai-labs/hibiki*.

We address the problem of extending a pre-trained large language model to a new domain that was not seen during training. Standard techniques, such as fine-tuning or low-rank adaptation (LoRA) are successful at domain adaptation, but do not formally add capacity to the model. This often leads to a trade-off, between performing well on the new domain vs. degrading performance on the original domain. Here, we propose to revisit and improve adapters to extend LLMs. Our paper analyzes this extension problem from three angles: data, architecture and training procedure, which are advantageously considered jointly. The resulting method, called neutral residues, modifies adapters in a way that leads to each new residual block to output near-zeros on the original domain. This solution leads to strong results when adapting a state-of-the-art model originally trained on English to a new language. Neutral residues significantly outperforms competing approaches such as fine-tuning, LoRA or vanilla adapters in terms of the trade-off between learning the new language and not forgetting English.

Large language models have shown impressive few-shot results on a wide range of tasks. However, when knowledge is key for such results, as is the case for tasks such as question answering and fact checking, massive parameter counts to store knowledge seem to be needed. Retrieval-augmented models are known to excel at knowledge intensive tasks without the need for as many parameters, but it is unclear whether they work in few-shot settings. In this work we present Atlas, a carefully designed and pre-trained retrieval-augmented language model able to learn knowledge intensive tasks with very few training examples. We perform evaluations on a wide range of tasks, including MMLU, KILT and Natural Questions, and study the impact of the content of the document index, showing that it can easily be updated. Notably, Atlas reaches over 42% accuracy on Natural Questions using only 64 examples, outperforming a 540B parameter model by 3% despite having 50x fewer parameters.

We present ResMLP, an architecture built entirely upon multi-layer perceptrons for image classification. It is a simple residual network that alternates (i) a linear layer in which image patches interact, independently and identically across channels, and (ii) a two-layer feed-forward network in which channels interact independently per patch. When trained with a modern training strategy using heavy data-augmentation and optionally distillation, it attains surprisingly good accuracy/complexity trade-offs on ImageNet. We also train ResMLP models in a self-supervised setup, to further remove priors from employing a labelled dataset. Finally, by adapting our model to machine translation we achieve surprisingly good results. We share pre-trained models and our code based on the Timm library.

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.

Holger Schwenk, Guillaume Wenzek, Sergey Edunov, Edouard Grave, Armand Joulin, Angela Fan. Proceedings of the 59th Annual Meeting of the Association for Computational Linguistics and the 11th International Joint Conference on Natural Language Processing (Volume 1: Long Papers). 2021.

Generative models for open domain question answering have proven to be competitive, without resorting to external knowledge. While promising, this approach requires to use models with billions of parameters, which are expensive to train and query. In this paper, we investigate how much these models can benefit from retrieving text passages, potentially containing evidence. We obtain state-of-the-art results on the Natural Questions and TriviaQA open benchmarks. Interestingly, we observe that the performance of this method significantly improves when increasing the number of retrieved passages. This is evidence that generative models are good at aggregating and combining evidence from multiple passages.

Jingfei Du, Edouard Grave, Beliz Gunel, Vishrav Chaudhary, Onur Celebi, Michael Auli, Veselin Stoyanov, Alexis Conneau. Proceedings of the 2021 Conference of the North American Chapter of the Association for Computational Linguistics: Human Language Technologies. 2021.

Alexis Conneau, Kartikay Khandelwal, Naman Goyal, Vishrav Chaudhary, Guillaume Wenzek, Francisco Guzmán, Edouard Grave, Myle Ott, Luke Zettlemoyer, Veselin Stoyanov. Proceedings of the 58th Annual Meeting of the Association for Computational Linguistics. 2020.

Pre-training text representations have led to significant improvements in many areas of natural language processing. The quality of these models benefits greatly from the size of the pretraining corpora as long as its quality is preserved. In this paper, we describe an automatic pipeline to extract massive high-quality monolingual datasets from Common Crawl for a variety of languages. Our pipeline follows the data processing introduced in fastText (Mikolov et al., 2017; Grave et al., 2018), that deduplicates documents and identifies their language. We augment this pipeline with a filtering step to select documents that are close to high quality corpora like Wikipedia.

In this paper, we study the problem of hybrid language modeling, that is using models which can predict both characters and larger units such as character ngrams or words.Using such models, multiple potential segmentations usually exist for a given string, for example one using words and one using characters only.Thus, the probability of a string is the sum of the probabilities of all the possible segmentations.Here, we show how it is possible to marginalize over the segmentations efficiently, in order to compute the true probability of a sequence.We apply our technique on three datasets, comprising seven languages, showing improvements over a strong character level language model.

We propose a novel self-attention mechanism that can learn its optimal attention span.This allows us to extend significantly the maximum context size used in Transformer, while maintaining control over their memory footprint and computational time.We show the effectiveness of our approach on the task of character level language modeling, where we achieve state-of-the-art performances on text8 and enwiki8 by using a maximum context of 8k characters.

Alex Wang, Jan Hula, Patrick Xia, Raghavendra Pappagari, R. Thomas McCoy, Roma Patel, Najoung Kim, Ian Tenney, Yinghui Huang, Katherin Yu, Shuning Jin, Berlin Chen, Benjamin Van Durme, Edouard Grave, Ellie Pavlick, Samuel R. Bowman. Proceedings of the 57th Annual Meeting of the Association for Computational Linguistics. 2019.

We consider the task of aligning two sets of points in high dimension, which has many applications in natural language processing and computer vision. As an example, it was recently shown that it is possible to infer a bilingual lexicon, without supervised data, by aligning word embeddings trained on monolingual data. These recent advances are based on adversarial training to learn the mapping between the two embeddings. In this paper, we propose to use an alternative formulation, based on the joint estimation of an orthogonal matrix and a permutation matrix. While this problem is not convex, we propose to initialize our optimization algorithm by using a convex relaxation, traditionally considered for the graph isomorphism problem. We propose a stochastic algorithm to minimize our cost function on large scale problems. Finally, we evaluate our method on the problem of unsupervised word translation, by aligning word embeddings trained on monolingual data. On this task, our method obtains state of the art results, while requiring less computational resources than competing approaches.

Paula Czarnowska, Sebastian Ruder, Edouard Grave, Ryan Cotterell, Ann Copestake. Proceedings of the 2019 Conference on Empirical Methods in Natural Language Processing and the 9th International Joint Conference on Natural Language Processing (EMNLP-IJCNLP). 2019.

While the incipient internet was largely text-based, the modern digital world is becoming increasingly multi-modal. Here, we examine multi-modal classification where one modality is discrete, e.g. text, and the other is continuous, e.g. visual representations transferred from a convolutional neural network. In particular, we focus on scenarios where we have to be able to classify large quantities of data quickly. We investigate various methods for performing multi-modal fusion and analyze their trade-offs in terms of classification accuracy and computational efficiency. Our findings indicate that the inclusion of continuous information improves performance over text-only on a range of multi-modal classification tasks, even with simple fusion methods. In addition, we experiment with discretizing the continuous features in order to speed up and simplify the fusion process even further. Our results show that fusion with discretized features outperforms text-only classification, at a fraction of the computational cost of full multi-modal fusion, with the additional benefit of improved interpretability.

Continuous word representations learned separately on distinct languages can be aligned so that their words become comparable in a common space.Existing works typically solve a quadratic problem to learn a orthogonal matrix aligning a bilingual lexicon, and use a retrieval criterion for inference.In this paper, we propose an unified formulation that directly optimizes a retrieval criterion in an end-to-end fashion.Our experiments on standard benchmarks show that our approach outperforms the state of the art on word translation, with the biggest improvements observed for distant language pairs such as English-Chinese.

Kristina Gulordava, Piotr Bojanowski, Edouard Grave, Tal Linzen, Marco Baroni. Proceedings of the 2018 Conference of the North American Chapter of the Association for Computational Linguistics: Human Language Technologies, Volume 1 (Long Papers). 2018.

We introduce Parseval networks, a form of deep neural networks in which the Lipschitz constant of linear, convolutional and aggregation layers is constrained to be smaller than $1$. Parseval networks are empirically and theoretically motivated by an analysis of the robustness of the predictions made by deep neural networks when their input is subject to an adversarial perturbation. The most important feature of Parseval networks is to maintain weight matrices of linear and convolutional layers to be (approximately) Parseval tight frames, which are extensions of orthogonal matrices to non-square matrices. We describe how these constraints can be maintained efficiently during SGD. We show that Parseval networks match the state-of-the-art regarding accuracy on CIFAR-10/100 and Street View House Numbers (SVHN), while being more robust than their vanilla counterpart against adversarial examples. Incidentally, Parseval networks also tend to train faster and make a better usage of the full capacity of the networks.

We propose an approximate strategy to efficiently train neural network based language models over very large vocabularies. Our approach, called adaptive softmax, circumvents the linear dependency on the vocabulary size by exploiting the unbalanced word distribution to form clusters that explicitly minimize the expectation of computation time. Our approach further reduces the computational cost by exploiting the specificities of modern architectures and matrix-matrix vector operations, making it particularly suited for graphical processing units. Our experiments carried out on standard benchmarks, such as EuroParl and One Billion Word, show that our approach brings a large gain in efficiency over standard approximations while achieving an accuracy close to that of the full softmax. The code of our method is available at https://github.com/facebookresearch/adaptive-softmax.

Recently, continuous cache models were proposed as extensions to recurrent neural network language models, to adapt their predictions to local changes in the data distribution. These models only capture the local context, of up to a few thousands tokens. In this paper, we propose an extension of continuous cache models, which can scale to larger contexts. In particular, we use a large scale non-parametric memory component that stores all the hidden activations seen in the past. We leverage recent advances in approximate nearest neighbor search and quantization algorithms to store millions of representations while searching them efficiently. We conduct extensive experiments showing that our approach significantly improves the perplexity of pre-trained language models on new distributions, and can scale efficiently to much larger contexts than previously proposed local cache models.

Identifying topics of discussions in online health communities (OHC) is critical to various information extraction applications, but can be difficult because topics of OHC content are usually heterogeneous and domain-dependent. In this paper, we provide a multi-class schema, an annotated dataset, and supervised classifiers based on convolutional neural network (CNN) and other models for the task of classifying discussion topics. We apply the CNN classifier to the most popular breast cancer online community, and carry out cross-sectional and longitudinal analyses to show topic distributions and topic dynamics throughout members' participation. Our experimental results suggest that CNN outperforms other classifiers in the task of topic classification and identify several patterns and trajectories. For example, although members discuss mainly disease-related topics, their interest may change through time and vary with their disease severities.

Continuous word representations, trained on large unlabeled corpora are useful for many natural language processing tasks. Popular models that learn such representations ignore the morphology of words, by assigning a distinct vector to each word. This is a limitation, especially for languages with large vocabularies and many rare words. In this paper, we propose a new approach based on the skipgram model, where each word is represented as a bag of character n-grams. A vector representation is associated to each character n-gram; words being represented as the sum of these representations. Our method is fast, allowing to train models on large corpora quickly and allows us to compute word representations for words that did not appear in the training data. We evaluate our word representations on nine different languages, both on word similarity and analogy tasks. By comparing to recently proposed morphological word representations, we show that our vectors achieve state-of-the-art performance on these tasks.

Édouard Grave, Noémie Elhadad. 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.

Suppose that we are given a set of videos, along with natural language descriptions in the form of multiple sentences (e.g., manual annotations, movie scripts, sport summaries etc.), and that these sentences appear in the same temporal order as their visual counterparts. We propose in this paper a method for aligning the two modalities, i.e., automatically providing a time (frame) stamp for every sentence. Given vectorial features for both video and text, this can be cast as a temporal assignment problem, with an implicit linear mapping between the two feature modalities. We formulate this problem as an integer quadratic program, and solve its continuous convex relaxation using an efficient conditional gradient algorithm. Several rounding procedures are proposed to construct the final integer solution. After demonstrating significant improvements over the state of the art on the related task of aligning video with symbolic labels [7], we evaluate our method on a challenging dataset of videos with associated textual descriptions [37], and explore bag-of-words and continuous representations for text.

A promising approach to relation extrac-tion, called weak or distant supervision, exploits an existing database of facts as training data, by aligning it to an unla-beled collection of text documents. Using this approach, the task of relation extrac-tion can easily be scaled to hundreds of different relationships. However, distant supervision leads to a challenging multi-ple instance, multiple label learning prob-lem. Most of the proposed solutions to this problem are based on non-convex formu-lations, and are thus prone to local min-ima. In this article, we propose a new approach to the problem of weakly su-pervised relation extraction, based on dis-criminative clustering and leading to a convex formulation. We demonstrate that our approach outperforms state-of-the-art methods on the challenging dataset intro-duced by Riedel et al. (2010). 1

Using the $\ell_1$-norm to regularize the estimation of the parameter vector of a linear model leads to an unstable estimator when covariates are highly correlated. In this paper, we introduce a new penalty function which takes into account the correlation of the design matrix to stabilize the estimation. This norm, called the trace Lasso, uses the trace norm of the selected covariates, which is a convex surrogate of their rank, as the criterion of model complexity. We analyze the properties of our norm, describe an optimization algorithm based on reweighted least-squares, and illustrate the behavior of this norm on synthetic data, showing that it is more adapted to strong correlations than competing methods such as the elastic net.