Amélie Royer

Alexandre Défossez, Laurent Mazaré, Manu Orsini et al.

We introduce Moshi, a speech-text foundation model and full-duplex spoken dialogue framework. Current systems for spoken dialogue rely on pipelines of independent components, namely voice activity detection, speech recognition, textual dialogue and text-to-speech. Such frameworks cannot emulate the experience of real conversations. First, their complexity induces a latency of several seconds between interactions. Second, text being the intermediate modality for dialogue, non-linguistic information that modifies meaning -- such as emotion or non-speech sounds -- is lost in the interaction. Finally, they rely on a segmentation into speaker turns, which does not take into account overlapping speech, interruptions and interjections. Moshi solves these independent issues altogether by casting spoken dialogue as speech-to-speech generation. Starting from a text language model backbone, Moshi generates speech as tokens from the residual quantizer of a neural audio codec, while modeling separately its own speech and that of the user into parallel streams. This allows for the removal of explicit speaker turns, and the modeling of arbitrary conversational dynamics. We moreover extend the hierarchical semantic-to-acoustic token generation of previous work to first predict time-aligned text tokens as a prefix to audio tokens. Not only this "Inner Monologue" method significantly improves the linguistic quality of generated speech, but we also illustrate how it can provide streaming speech recognition and text-to-speech. Our resulting model is the first real-time full-duplex spoken large language model, with a theoretical latency of 160ms, 200ms in practice, and is available at https://github.com/kyutai-labs/moshi.

Amelie Royer, Tijmen Blankevoort, Babak Ehteshami Bejnordi

Training a single model on multiple input domains and/or output tasks allows for compressing information from multiple sources into a unified backbone hence improves model efficiency. It also enables potential positive knowledge transfer across tasks/domains, leading to improved accuracy and data-efficient training. However, optimizing such networks is a challenge, in particular due to discrepancies between the different tasks or domains: Despite several hypotheses and solutions proposed over the years, recent work has shown that uniform scalarization training, i.e., simply minimizing the average of the task losses, yields on-par performance with more costly SotA optimization methods. This raises the issue of how well we understand the training dynamics of multi-task and multi-domain networks. In this work, we first devise a large-scale unified analysis of multi-domain and multi-task learning to better understand the dynamics of scalarization across varied task/domain combinations and model sizes. Following these insights, we then propose to leverage population-based training to efficiently search for the optimal scalarization weights when dealing with a large number of tasks or domains.

Jakob Drachmann Havtorn, Amelie Royer, Tijmen Blankevoort et al.

The input tokens to Vision Transformers carry little semantic meaning as they are defined as regular equal-sized patches of the input image, regardless of its content. However, processing uniform background areas of an image should not necessitate as much compute as dense, cluttered areas. To address this issue, we propose a dynamic mixed-scale tokenization scheme for ViT, MSViT. Our method introduces a conditional gating mechanism that selects the optimal token scale for every image region, such that the number of tokens is dynamically determined per input. In addition, to enhance the conditional behavior of the gate during training, we introduce a novel generalization of the batch-shaping loss. We show that our gating module is able to learn meaningful semantics despite operating locally at the coarse patch-level. The proposed gating module is lightweight, agnostic to the choice of transformer backbone, and trained within a few epochs with little training overhead. Furthermore, in contrast to token pruning, MSViT does not lose information about the input, thus can be readily applied for dense tasks. We validate MSViT on the tasks of classification and segmentation where it leads to improved accuracy-complexity trade-off.

Amelie Royer, Ilia Karmanov, Andrii Skliar et al.

Mixture of Experts (MoE) are rising in popularity as a means to train extremely large-scale models, yet allowing for a reasonable computational cost at inference time. Recent state-of-the-art approaches usually assume a large number of experts, and require training all experts jointly, which often lead to training instabilities such as the router collapsing In contrast, in this work, we propose to revisit the simple single-gate MoE, which allows for more practical training. Key to our work are (i) a base model branch acting both as an early-exit and an ensembling regularization scheme, (ii) a simple and efficient asynchronous training pipeline without router collapse issues, and finally (iii) a per-sample clustering-based initialization. We show experimentally that the proposed model obtains efficiency-to-accuracy trade-offs comparable with other more complex MoE, and outperforms non-mixture baselines. This showcases the merits of even a simple single-gate MoE, and motivates further exploration in this area.

Moritz Böhle, Amélie Royer, Juliette Marrie et al.

Vision-language models (VLMs) are commonly trained by inserting image tokens from a pretrained vision encoder into the textual stream of a language model. This allows text and image information to fully attend to one another within the model, but becomes extremely costly for high-resolution images, long conversations, or streaming videos, both in memory and compute. VLMs leveraging cross-attention are an efficient alternative to token insertion but exhibit a clear performance gap, in particular on tasks involving fine-grained visual details. We find that a key to improving such models is to also enable local text-to-text interaction in the dedicated cross-attention layers. Building on this, we propose CASA, Cross-Attention via Self-Attention, a simple and efficient paradigm which substantially reduces the gap with full token insertion on common image understanding benchmarks, while enjoying the same scalability as cross-attention models when applied to long-context multimodal tasks such as streaming video captioning. For samples and code, please see our project page at https://kyutai.org/casa .

Benjamin Bergner, Andrii Skliar, Amelie Royer et al.

Large language models (LLMs) have become ubiquitous in practice and are widely used for generation tasks such as translation, summarization and instruction following. However, their enormous size and reliance on autoregressive decoding increase deployment costs and complicate their use in latency-critical applications. In this work, we propose a hybrid approach that combines language models of different sizes to increase the efficiency of autoregressive decoding while maintaining high performance. Our method utilizes a pretrained frozen LLM that encodes all prompt tokens once in parallel, and uses the resulting representations to condition and guide a small language model (SLM), which then generates the response more efficiently. We investigate the combination of encoder-decoder LLMs with both encoder-decoder and decoder-only SLMs from different model families and only require fine-tuning of the SLM. Experiments with various benchmarks show substantial speedups of up to $4\times$, with minor performance penalties of $1-2\%$ for translation and summarization tasks compared to the LLM.

Babak Ehteshami Bejnordi, Gaurav Kumar, Amelie Royer et al.

Jointly learning multiple tasks with a unified model can improve accuracy and data efficiency, but it faces the challenge of task interference, where optimizing one task objective may inadvertently compromise the performance of another. A solution to mitigate this issue is to allocate task-specific parameters, free from interference, on top of shared features. However, manually designing such architectures is cumbersome, as practitioners need to balance between the overall performance across all tasks and the higher computational cost induced by the newly added parameters. In this work, we propose \textit{InterroGate}, a novel multi-task learning (MTL) architecture designed to mitigate task interference while optimizing inference computational efficiency. We employ a learnable gating mechanism to automatically balance the shared and task-specific representations while preserving the performance of all tasks. Crucially, the patterns of parameter sharing and specialization dynamically learned during training, become fixed at inference, resulting in a static, optimized MTL architecture. Through extensive empirical evaluations, we demonstrate SoTA results on three MTL benchmarks using convolutional as well as transformer-based backbones on CelebA, NYUD-v2, and PASCAL-Context.

Amélie Royer, Moritz Böhle, Gabriel de Marmiesse et al.

The recent successes of Vision-Language models raise the question of how to equivalently imbue a pretrained speech model with vision understanding, an important milestone towards building a multimodal speech model able to freely converse about images. Building such a conversational Vision-Speech model brings its unique challenges: (i) paired image-speech datasets are much scarcer than their image-text counterparts, (ii) ensuring real-time latency at inference is crucial thus bringing compute and memory constraints, and (iii) the model should preserve prosodic features (e.g., speaker tone) which cannot be inferred from text alone. In this work, we introduce MoshiVis, augmenting a recent dialogue speech LLM, Moshi, with visual inputs through lightweight adaptation modules. An additional dynamic gating mechanism enables the model to more easily switch between the visual inputs and unrelated conversation topics. To reduce training costs, we design a simple one-stage, parameter-efficient fine-tuning pipeline in which we leverage a mixture of image-text (i.e., "speechless") and image-speech samples. We evaluate the model on downstream visual understanding tasks with both audio and text prompts, and report qualitative samples of interactions with MoshiVis. Our inference code will be made available, as well as the image-speech data used for audio evaluation.

Lucas Beyer, Xiaohua Zhai, Amélie Royer et al.

There is a growing discrepancy in computer vision between large-scale models that achieve state-of-the-art performance and models that are affordable in practical applications. In this paper we address this issue and significantly bridge the gap between these two types of models. Throughout our empirical investigation we do not aim to necessarily propose a new method, but strive to identify a robust and effective recipe for making state-of-the-art large scale models affordable in practice. We demonstrate that, when performed correctly, knowledge distillation can be a powerful tool for reducing the size of large models without compromising their performance. In particular, we uncover that there are certain implicit design choices, which may drastically affect the effectiveness of distillation. Our key contribution is the explicit identification of these design choices, which were not previously articulated in the literature. We back up our findings by a comprehensive empirical study, demonstrate compelling results on a wide range of vision datasets and, in particular, obtain a state-of-the-art ResNet-50 model for ImageNet, which achieves 82.8% top-1 accuracy.

Amelie Royer, Christoph H. Lampert

Fine-tuning is a popular way of exploiting knowledge contained in a pre-trained convolutional network for a new visual recognition task. However, the orthogonal setting of transferring knowledge from a pretrained network to a visually different yet semantically close source is rarely considered: This commonly happens with real-life data, which is not necessarily as clean as the training source (noise, geometric transformations, different modalities, etc.). To tackle such scenarios, we introduce a new, generalized form of fine-tuning, called flex-tuning, in which any individual unit (e.g. layer) of a network can be tuned, and the most promising one is chosen automatically. In order to make the method appealing for practical use, we propose two lightweight and faster selection procedures that prove to be good approximations in practice. We study these selection criteria empirically across a variety of domain shifts and data scarcity scenarios, and show that fine-tuning individual units, despite its simplicity, yields very good results as an adaptation technique. As it turns out, in contrast to common practice, rather than the last fully-connected unit it is best to tune an intermediate or early one in many domain-shift scenarios, which is accurately detected by flex-tuning.

Amelie Royer, Christoph H. Lampert

State-of-the-art detection systems are generally evaluated on their ability to exhaustively retrieve objects densely distributed in the image, across a wide variety of appearances and semantic categories. Orthogonal to this, many real-life object detection applications, for example in remote sensing, instead require dealing with large images that contain only a few small objects of a single class, scattered heterogeneously across the space. In addition, they are often subject to strict computational constraints, such as limited battery capacity and computing power. To tackle these more practical scenarios, we propose a novel flexible detection scheme that efficiently adapts to variable object sizes and densities: We rely on a sequence of detection stages, each of which has the ability to predict groups of objects as well as individuals. Similar to a detection cascade, this multi-stage architecture spares computational effort by discarding large irrelevant regions of the image early during the detection process. The ability to group objects provides further computational and memory savings, as it allows working with lower image resolutions in early stages, where groups are more easily detected than individuals, as they are more salient. We report experimental results on two aerial image datasets, and show that the proposed method is as accurate yet computationally more efficient than standard single-shot detectors, consistently across three different backbone architectures.

Amélie Royer, Konstantinos Bousmalis, Stephan Gouws et al.

Style transfer usually refers to the task of applying color and texture information from a specific style image to a given content image while preserving the structure of the latter. Here we tackle the more generic problem of semantic style transfer: given two unpaired collections of images, we aim to learn a mapping between the corpus-level style of each collection, while preserving semantic content shared across the two domains. We introduce XGAN ("Cross-GAN"), a dual adversarial autoencoder, which captures a shared representation of the common domain semantic content in an unsupervised way, while jointly learning the domain-to-domain image translations in both directions. We exploit ideas from the domain adaptation literature and define a semantic consistency loss which encourages the model to preserve semantics in the learned embedding space. We report promising qualitative results for the task of face-to-cartoon translation. The cartoon dataset, CartoonSet, we collected for this purpose is publicly available at google.github.io/cartoonset/ as a new benchmark for semantic style transfer.

Amelie Royer, Alexander Kolesnikov, Christoph H. Lampert

We develop a probabilistic technique for colorizing grayscale natural images. In light of the intrinsic uncertainty of this task, the proposed probabilistic framework has numerous desirable properties. In particular, our model is able to produce multiple plausible and vivid colorizations for a given grayscale image and is one of the first colorization models to provide a proper stochastic sampling scheme. Moreover, our training procedure is supported by a rigorous theoretical framework that does not require any ad hoc heuristics and allows for efficient modeling and learning of the joint pixel color distribution. We demonstrate strong quantitative and qualitative experimental results on the CIFAR-10 dataset and the challenging ILSVRC 2012 dataset.