Hervé Jégou

Pierre Fernandez, Guillaume Couairon, Hervé Jégou et al. · meta-ai

Generative image modeling enables a wide range of applications but raises ethical concerns about responsible deployment. This paper introduces an active strategy combining image watermarking and Latent Diffusion Models. The goal is for all generated images to conceal an invisible watermark allowing for future detection and/or identification. The method quickly fine-tunes the latent decoder of the image generator, conditioned on a binary signature. A pre-trained watermark extractor recovers the hidden signature from any generated image and a statistical test then determines whether it comes from the generative model. We evaluate the invisibility and robustness of the watermarks on a variety of generation tasks, showing that Stable Signature works even after the images are modified. For instance, it detects the origin of an image generated from a text prompt, then cropped to keep $10\%$ of the content, with $90$+$\%$ accuracy at a false positive rate below 10$^{-6}$.

Matthew J. Muckley, Alaaeldin El-Nouby, Karen Ullrich et al. · meta-ai

Lossy image compression aims to represent images in as few bits as possible while maintaining fidelity to the original. Theoretical results indicate that optimizing distortion metrics such as PSNR or MS-SSIM necessarily leads to a discrepancy in the statistics of original images from those of reconstructions, in particular at low bitrates, often manifested by the blurring of the compressed images. Previous work has leveraged adversarial discriminators to improve statistical fidelity. Yet these binary discriminators adopted from generative modeling tasks may not be ideal for image compression. In this paper, we introduce a non-binary discriminator that is conditioned on quantized local image representations obtained via VQ-VAE autoencoders. Our evaluations on the CLIC2020, DIV2K and Kodak datasets show that our discriminator is more effective for jointly optimizing distortion (e.g., PSNR) and statistical fidelity (e.g., FID) than the PatchGAN of the state-of-the-art HiFiC model. On CLIC2020, we obtain the same FID as HiFiC with 30-40\% fewer bits.

Pierre Fernandez, Matthijs Douze, Hervé Jégou et al. · meta-ai

Image copy detection and retrieval from large databases leverage two components. First, a neural network maps an image to a vector representation, that is relatively robust to various transformations of the image. Second, an efficient but approximate similarity search algorithm trades scalability (size and speed) against quality of the search, thereby introducing a source of error. This paper improves the robustness of image copy detection with active indexing, that optimizes the interplay of these two components. We reduce the quantization loss of a given image representation by making imperceptible changes to the image before its release. The loss is back-propagated through the deep neural network back to the image, under perceptual constraints. These modifications make the image more retrievable. Our experiments show that the retrieval and copy detection of activated images is significantly improved. For instance, activation improves by $+40\%$ the Recall1@1 on various image transformations, and for several popular indexing structures based on product quantization and locality sensitivity hashing.

Hugo Touvron, Matthieu Cord, Hervé Jégou

A Vision Transformer (ViT) is a simple neural architecture amenable to serve several computer vision tasks. It has limited built-in architectural priors, in contrast to more recent architectures that incorporate priors either about the input data or of specific tasks. Recent works show that ViTs benefit from self-supervised pre-training, in particular BerT-like pre-training like BeiT. In this paper, we revisit the supervised training of ViTs. Our procedure builds upon and simplifies a recipe introduced for training ResNet-50. It includes a new simple data-augmentation procedure with only 3 augmentations, closer to the practice in self-supervised learning. Our evaluations on Image classification (ImageNet-1k with and without pre-training on ImageNet-21k), transfer learning and semantic segmentation show that our procedure outperforms by a large margin previous fully supervised training recipes for ViT. It also reveals that the performance of our ViT trained with supervision is comparable to that of more recent architectures. Our results could serve as better baselines for recent self-supervised approaches demonstrated on ViT.

Alaaeldin El-Nouby, Matthew J. Muckley, Karen Ullrich et al. · meta-ai

Recent neural compression methods have been based on the popular hyperprior framework. It relies on Scalar Quantization and offers a very strong compression performance. This contrasts from recent advances in image generation and representation learning, where Vector Quantization is more commonly employed. In this work, we attempt to bring these lines of research closer by revisiting vector quantization for image compression. We build upon the VQ-VAE framework and introduce several modifications. First, we replace the vanilla vector quantizer by a product quantizer. This intermediate solution between vector and scalar quantization allows for a much wider set of rate-distortion points: It implicitly defines high-quality quantizers that would otherwise require intractably large codebooks. Second, inspired by the success of Masked Image Modeling (MIM) in the context of self-supervised learning and generative image models, we propose a novel conditional entropy model which improves entropy coding by modelling the co-dependencies of the quantized latent codes. The resulting PQ-MIM model is surprisingly effective: its compression performance on par with recent hyperprior methods. It also outperforms HiFiC in terms of FID and KID metrics when optimized with perceptual losses (e.g. adversarial). Finally, since PQ-MIM is compatible with image generation frameworks, we show qualitatively that it can operate under a hybrid mode between compression and generation, with no further training or finetuning. As a result, we explore the extreme compression regime where an image is compressed into 200 bytes, i.e., less than a tweet.

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.

Hugo Touvron, Matthieu Cord, Alaaeldin El-Nouby et al.

After their initial success in natural language processing, transformer architectures have rapidly gained traction in computer vision, providing state-of-the-art results for tasks such as image classification, detection, segmentation, and video analysis. We offer three insights based on simple and easy to implement variants of vision transformers. (1) The residual layers of vision transformers, which are usually processed sequentially, can to some extent be processed efficiently in parallel without noticeably affecting the accuracy. (2) Fine-tuning the weights of the attention layers is sufficient to adapt vision transformers to a higher resolution and to other classification tasks. This saves compute, reduces the peak memory consumption at fine-tuning time, and allows sharing the majority of weights across tasks. (3) Adding MLP-based patch pre-processing layers improves Bert-like self-supervised training based on patch masking. We evaluate the impact of these design choices using the ImageNet-1k dataset, and confirm our findings on the ImageNet-v2 test set. Transfer performance is measured across six smaller datasets.

Hugo Touvron, Matthieu Cord, Maxime Oquab et al.

We introduce submodel co-training, a regularization method related to co-training, self-distillation and stochastic depth. Given a neural network to be trained, for each sample we implicitly instantiate two altered networks, ``submodels'', with stochastic depth: we activate only a subset of the layers. Each network serves as a soft teacher to the other, by providing a loss that complements the regular loss provided by the one-hot label. Our approach, dubbed cosub, uses a single set of weights, and does not involve a pre-trained external model or temporal averaging. Experimentally, we show that submodel co-training is effective to train backbones for recognition tasks such as image classification and semantic segmentation. Our approach is compatible with multiple architectures, including RegNet, ViT, PiT, XCiT, Swin and ConvNext. Our training strategy improves their results in comparable settings. For instance, a ViT-B pretrained with cosub on ImageNet-21k obtains 87.4% top-1 acc. @448 on ImageNet-val.

Huy V. Vo, Vasil Khalidov, Timothée Darcet et al.

Self-supervised features are the cornerstone of modern machine learning systems. They are typically pre-trained on data collections whose construction and curation typically require extensive human effort. This manual process has some limitations similar to those encountered in supervised learning, e.g., the crowd-sourced selection of data is costly and time-consuming, preventing scaling the dataset size. In this work, we consider the problem of automatic curation of high-quality datasets for self-supervised pre-training. We posit that such datasets should be large, diverse and balanced, and propose a clustering-based approach for building ones satisfying all these criteria. Our method involves successive and hierarchical applications of $k$-means on a large and diverse data repository to obtain clusters that distribute uniformly among data concepts, followed by a hierarchical, balanced sampling step from these clusters. Extensive experiments on three different data domains including web-based images, satellite images and text show that features trained on our automatically curated datasets outperform those trained on uncurated data while being on par or better than ones trained on manually curated data. Code is available at https://github.com/facebookresearch/ssl-data-curation.

Alexis Marouani, Oriane Siméoni, Hervé Jégou et al.

Vision Transformers have emerged as powerful, scalable and versatile representation learners. To capture both global and local features, a learnable [CLS] class token is typically prepended to the input sequence of patch tokens. Despite their distinct nature, both token types are processed identically throughout the model. In this work, we investigate the friction between global and local feature learning under different pre-training strategies by analyzing the interactions between class and patch tokens. Our analysis reveals that standard normalization layers introduce an implicit differentiation between these token types. Building on this insight, we propose specialized processing paths that selectively disentangle the computational flow of class and patch tokens, particularly within normalization layers and early query-key-value projections. This targeted specialization leads to significantly improved patch representation quality for dense prediction tasks. Our experiments demonstrate segmentation performance gains of over 2 mIoU points on standard benchmarks, while maintaining strong classification accuracy. The proposed modifications introduce only an 8% increase in parameters, with no additional computational overhead. Through comprehensive ablations, we provide insights into which architectural components benefit most from specialization and how our approach generalizes across model scales and learning frameworks.

Gergely Szilvasy, Manuel Faysse, Maria Lomeli et al.

Under modern test-time compute and agentic paradigms, language models process ever-longer sequences. Efficient text generation with transformer architectures is increasingly constrained by the Key-Value cache memory footprint and bandwidth. To address this limitation, we introduce Self-Pruned Key-Value Attention (SP-KV), a mechanism designed to predict future KV utility in order to reduce the size of the long-term KV cache. This strategy operates at a fine granularity: a lightweight utility predictor scores each key-value pair, and while recent KVs are always available via a local window, older pairs are written in the cache and used in global attention only if their predicted utility surpasses a given threshold. The LLM and the utility predictor are trained jointly end-to-end exclusively through next-token prediction loss, and are adapted from pretrained LLM checkpoints. Rather than enforcing a fixed compression ratio, SP-KV performs dynamic sparsification: the mechanism adapts to the input and typically reduces the KV cache size by a factor of $3$ to $10\times$, longer sequences often being more compressible. This leads to vast improvements in memory usage and decoding speed, with little to no degradation of validation loss nor performance on a broad set of downstream tasks. Beyond serving as an effective KV-cache reduction mechanism, our method reveals structured layer- and head-specific sparsity patterns that we can use to guide the design of hybrid local-global attention architectures.

Oriane Siméoni, Huy V. Vo, Maximilian Seitzer et al.

Self-supervised learning holds the promise of eliminating the need for manual data annotation, enabling models to scale effortlessly to massive datasets and larger architectures. By not being tailored to specific tasks or domains, this training paradigm has the potential to learn visual representations from diverse sources, ranging from natural to aerial images -- using a single algorithm. This technical report introduces DINOv3, a major milestone toward realizing this vision by leveraging simple yet effective strategies. First, we leverage the benefit of scaling both dataset and model size by careful data preparation, design, and optimization. Second, we introduce a new method called Gram anchoring, which effectively addresses the known yet unsolved issue of dense feature maps degrading during long training schedules. Finally, we apply post-hoc strategies that further enhance our models' flexibility with respect to resolution, model size, and alignment with text. As a result, we present a versatile vision foundation model that outperforms the specialized state of the art across a broad range of settings, without fine-tuning. DINOv3 produces high-quality dense features that achieve outstanding performance on various vision tasks, significantly surpassing previous self- and weakly-supervised foundation models. We also share the DINOv3 suite of vision models, designed to advance the state of the art on a wide spectrum of tasks and data by providing scalable solutions for diverse resource constraints and deployment scenarios.

Pierre Fernandez, Alexandre Sablayrolles, Teddy Furon et al.

We revisit watermarking techniques based on pre-trained deep networks, in the light of self-supervised approaches. We present a way to embed both marks and binary messages into their latent spaces, leveraging data augmentation at marking time. Our method can operate at any resolution and creates watermarks robust to a broad range of transformations (rotations, crops, JPEG, contrast, etc). It significantly outperforms the previous zero-bit methods, and its performance on multi-bit watermarking is on par with state-of-the-art encoder-decoder architectures trained end-to-end for watermarking. The code is available at github.com/facebookresearch/ssl_watermarking

Ross Wightman, Hugo Touvron, Hervé Jégou

The influential Residual Networks designed by He et al. remain the gold-standard architecture in numerous scientific publications. They typically serve as the default architecture in studies, or as baselines when new architectures are proposed. Yet there has been significant progress on best practices for training neural networks since the inception of the ResNet architecture in 2015. Novel optimization & data-augmentation have increased the effectiveness of the training recipes. In this paper, we re-evaluate the performance of the vanilla ResNet-50 when trained with a procedure that integrates such advances. We share competitive training settings and pre-trained models in the timm open-source library, with the hope that they will serve as better baselines for future work. For instance, with our more demanding training setting, a vanilla ResNet-50 reaches 80.4% top-1 accuracy at resolution 224x224 on ImageNet-val without extra data or distillation. We also report the performance achieved with popular models with our training procedure.

Ben Graham, Alaaeldin El-Nouby, Hugo Touvron et al.

We design a family of image classification architectures that optimize the trade-off between accuracy and efficiency in a high-speed regime. Our work exploits recent findings in attention-based architectures, which are competitive on highly parallel processing hardware. We revisit principles from the extensive literature on convolutional neural networks to apply them to transformers, in particular activation maps with decreasing resolutions. We also introduce the attention bias, a new way to integrate positional information in vision transformers. As a result, we propose LeVIT: a hybrid neural network for fast inference image classification. We consider different measures of efficiency on different hardware platforms, so as to best reflect a wide range of application scenarios. Our extensive experiments empirically validate our technical choices and show they are suitable to most architectures. Overall, LeViT significantly outperforms existing convnets and vision transformers with respect to the speed/accuracy tradeoff. For example, at 80% ImageNet top-1 accuracy, LeViT is 5 times faster than EfficientNet on CPU. We release the code at https://github.com/facebookresearch/LeViT

Jeff Johnson, Matthijs Douze, Hervé Jégou

Similarity search finds application in specialized database systems handling complex data such as images or videos, which are typically represented by high-dimensional features and require specific indexing structures. This paper tackles the problem of better utilizing GPUs for this task. While GPUs excel at data-parallel tasks, prior approaches are bottlenecked by algorithms that expose less parallelism, such as k-min selection, or make poor use of the memory hierarchy. We propose a design for k-selection that operates at up to 55% of theoretical peak performance, enabling a nearest neighbor implementation that is 8.5x faster than prior GPU state of the art. We apply it in different similarity search scenarios, by proposing optimized design for brute-force, approximate and compressed-domain search based on product quantization. In all these setups, we outperform the state of the art by large margins. Our implementation enables the construction of a high accuracy k-NN graph on 95 million images from the Yfcc100M dataset in 35 minutes, and of a graph connecting 1 billion vectors in less than 12 hours on 4 Maxwell Titan X GPUs. We have open-sourced our approach for the sake of comparison and reproducibility.

Edouard Grave, Armand Joulin, Moustapha Cissé et al.

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 time 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.

Pascale Fung, Yoram Bachrach, Asli Celikyilmaz et al.

This paper describes our research on AI agents embodied in visual, virtual or physical forms, enabling them to interact with both users and their environments. These agents, which include virtual avatars, wearable devices, and robots, are designed to perceive, learn and act within their surroundings, which makes them more similar to how humans learn and interact with the environments as compared to disembodied agents. We propose that the development of world models is central to reasoning and planning of embodied AI agents, allowing these agents to understand and predict their environment, to understand user intentions and social contexts, thereby enhancing their ability to perform complex tasks autonomously. World modeling encompasses the integration of multimodal perception, planning through reasoning for action and control, and memory to create a comprehensive understanding of the physical world. Beyond the physical world, we also propose to learn the mental world model of users to enable better human-agent collaboration.

Pierre-Emmanuel Mazaré, Gergely Szilvasy, Maria Lomeli et al.

Self-attention in transformer models is an incremental associative memory that maps key vectors to value vectors. One way to speed up self-attention is to employ GPU-compatible vector search algorithms based on standard partitioning methods such as k-means. However, such partitioning methods yield poor results in this context because (1) the keys and queries follow different distributions, and (2) the RoPE positional encoding hinders the bucket assignment. This paper introduces Saap (Self-Attention with Asymmetric Partitions), which overcomes these problems. It is an asymmetrical indexing technique that employs distinct partitions for keys and queries, thereby approximating self-attention with a data-adaptive sparsity pattern. It works on pretrained language models and only requires to train (offline) a small query classifier. On a long context Llama 3.1-8b model, with sequences ranging from 100k to 500k tokens, Saap typically reduces by a factor of 20 the fraction of memory that needs to be looked-up, which translates to a time saving of 60\% when compared to FlashAttention-v2.

Loïc Cabannes, Maximilian Beck, Gergely Szilvasy et al.

Recent works show that hybrid architectures combining sliding window softmax attention layers with linear recurrent neural network (RNN) layers outperform both of these architectures taken separately. However, the impact of the window length and the interplay between softmax attention and linear RNN layers remain under-studied. In this work, we introduce SWAX, a hybrid architecture consisting of sliding-window attention and xLSTM linear RNN layers. A counter-intuitive finding with SWAX is that larger sliding windows do not improve the long-context performance. In fact, short window attention encourages the model to better train the long-term memory of the xLSTM, by relying less on the softmax attention mechanism for long context-retrieval. The issue with small sliding windows is that they are detrimental for short-context tasks, which could be solved with information from moderately larger sliding windows otherwise. Therefore, we train SWAX by stochastically changing the sliding window size, forcing the model to leverage both a longer context window and the xLSTM memory. SWAX trained with stochastic window sizes significantly outperforms regular window attention both on short and long-context problems.

Maria Lomeli, Matthijs Douze, Gergely Szilvasy et al.

We introduce stochastic activations. This novel strategy randomly selects between several non-linear functions in the feed-forward layer of a large language model. In particular, we choose between SILU or RELU depending on a Bernoulli draw. This strategy circumvents the optimization problem associated with RELU, namely, the constant shape for negative inputs that prevents the gradient flow. We leverage this strategy in two ways: (1) We use stochastic activations during pre-training and fine-tune the model with RELU, which is used at inference time to provide sparse latent vectors. This reduces the inference FLOPs and translates into a significant speedup in the CPU. Interestingly, this leads to much better results than training from scratch with the RELU activation function. (2) We evaluate stochastic activations for generation. This strategy performs reasonably well: it is only slightly inferior to the best deterministic non-linearity, namely SILU combined with temperature scaling. This offers an alternative to existing strategies by providing a controlled way to increase the diversity of the generated text.

Matthijs Douze, Alexandr Guzhva, Chengqi Deng et al.

Vector databases typically manage large collections of embedding vectors. Currently, AI applications are growing rapidly, and so is the number of embeddings that need to be stored and indexed. The Faiss library is dedicated to vector similarity search, a core functionality of vector databases. Faiss is a toolkit of indexing methods and related primitives used to search, cluster, compress and transform vectors. This paper describes the trade-off space of vector search and the design principles of Faiss in terms of structure, approach to optimization and interfacing. We benchmark key features of the library and discuss a few selected applications to highlight its broad applicability.

Hugo Touvron, Matthieu Cord, Alaaeldin El-Nouby et al.

We show how to augment any convolutional network with an attention-based global map to achieve non-local reasoning. We replace the final average pooling by an attention-based aggregation layer akin to a single transformer block, that weights how the patches are involved in the classification decision. We plug this learned aggregation layer with a simplistic patch-based convolutional network parametrized by 2 parameters (width and depth). In contrast with a pyramidal design, this architecture family maintains the input patch resolution across all the layers. It yields surprisingly competitive trade-offs between accuracy and complexity, in particular in terms of memory consumption, as shown by our experiments on various computer vision tasks: object classification, image segmentation and detection.

Kenza Amara, Matthijs Douze, Alexandre Sablayrolles et al.

Modern approaches for fast retrieval of similar vectors on billion-scaled datasets rely on compressed-domain approaches such as binary sketches or product quantization. These methods minimize a certain loss, typically the mean squared error or other objective functions tailored to the retrieval problem. In this paper, we re-interpret popular methods such as binary hashing or product quantizers as auto-encoders, and point out that they implicitly make suboptimal assumptions on the form of the decoder. We design backward-compatible decoders that improve the reconstruction of the vectors from the same codes, which translates to a better performance in nearest neighbor search. Our method significantly improves over binary hashing methods or product quantization on popular benchmarks.

Hugo Touvron, Piotr Bojanowski, Mathilde Caron et al.

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.

Mathilde Caron, Hugo Touvron, Ishan Misra et al.

In this paper, we question if self-supervised learning provides new properties to Vision Transformer (ViT) that stand out compared to convolutional networks (convnets). Beyond the fact that adapting self-supervised methods to this architecture works particularly well, we make the following observations: first, self-supervised ViT features contain explicit information about the semantic segmentation of an image, which does not emerge as clearly with supervised ViTs, nor with convnets. Second, these features are also excellent k-NN classifiers, reaching 78.3% top-1 on ImageNet with a small ViT. Our study also underlines the importance of momentum encoder, multi-crop training, and the use of small patches with ViTs. We implement our findings into a simple self-supervised method, called DINO, which we interpret as a form of self-distillation with no labels. We show the synergy between DINO and ViTs by achieving 80.1% top-1 on ImageNet in linear evaluation with ViT-Base.

Chuan Guo, Alexandre Sablayrolles, Hervé Jégou et al.

We propose the first general-purpose gradient-based attack against transformer models. Instead of searching for a single adversarial example, we search for a distribution of adversarial examples parameterized by a continuous-valued matrix, hence enabling gradient-based optimization. We empirically demonstrate that our white-box attack attains state-of-the-art attack performance on a variety of natural language tasks. Furthermore, we show that a powerful black-box transfer attack, enabled by sampling from the adversarial distribution, matches or exceeds existing methods, while only requiring hard-label outputs.

Hugo Touvron, Matthieu Cord, Alexandre Sablayrolles et al.

Transformers have been recently adapted for large scale image classification, achieving high scores shaking up the long supremacy of convolutional neural networks. However the optimization of image transformers has been little studied so far. In this work, we build and optimize deeper transformer networks for image classification. In particular, we investigate the interplay of architecture and optimization of such dedicated transformers. We make two transformers architecture changes that significantly improve the accuracy of deep transformers. This leads us to produce models whose performance does not saturate early with more depth, for instance we obtain 86.5% top-1 accuracy on Imagenet when training with no external data, we thus attain the current SOTA with less FLOPs and parameters. Moreover, our best model establishes the new state of the art on Imagenet with Reassessed labels and Imagenet-V2 / match frequency, in the setting with no additional training data. We share our code and models.

Alaaeldin El-Nouby, Natalia Neverova, Ivan Laptev et al.

Transformers have shown outstanding results for natural language understanding and, more recently, for image classification. We here extend this work and propose a transformer-based approach for image retrieval: we adopt vision transformers for generating image descriptors and train the resulting model with a metric learning objective, which combines a contrastive loss with a differential entropy regularizer. Our results show consistent and significant improvements of transformers over convolution-based approaches. In particular, our method outperforms the state of the art on several public benchmarks for category-level retrieval, namely Stanford Online Product, In-Shop and CUB-200. Furthermore, our experiments on ROxford and RParis also show that, in comparable settings, transformers are competitive for particular object retrieval, especially in the regime of short vector representations and low-resolution images.

Hugo Touvron, Matthieu Cord, Matthijs Douze et al.

Recently, neural networks purely based on attention were shown to address image understanding tasks such as image classification. However, these visual transformers are pre-trained with hundreds of millions of images using an expensive infrastructure, thereby limiting their adoption. In this work, we produce a competitive convolution-free transformer by training on Imagenet only. We train them on a single computer in less than 3 days. Our reference vision transformer (86M parameters) achieves top-1 accuracy of 83.1% (single-crop evaluation) on ImageNet with no external data. More importantly, we introduce a teacher-student strategy specific to transformers. It relies on a distillation token ensuring that the student learns from the teacher through attention. We show the interest of this token-based distillation, especially when using a convnet as a teacher. This leads us to report results competitive with convnets for both Imagenet (where we obtain up to 85.2% accuracy) and when transferring to other tasks. We share our code and models.

Hugo Touvron, Alexandre Sablayrolles, Matthijs Douze et al.

This paper tackles the problem of learning a finer representation than the one provided by training labels. This enables fine-grained category retrieval of images in a collection annotated with coarse labels only. Our network is learned with a nearest-neighbor classifier objective, and an instance loss inspired by self-supervised learning. By jointly leveraging the coarse labels and the underlying fine-grained latent space, it significantly improves the accuracy of category-level retrieval methods. Our strategy outperforms all competing methods for retrieving or classifying images at a finer granularity than that available at train time. It also improves the accuracy for transfer learning tasks to fine-grained datasets, thereby establishing the new state of the art on five public benchmarks, like iNaturalist-2018.

Hugo Touvron, Matthijs Douze, Matthieu Cord et al.

We propose a simple architecture to address unpaired image-to-image translation tasks: style or class transfer, denoising, deblurring, deblocking, etc. We start from an image autoencoder architecture with fixed weights. For each task we learn a residual block operating in the latent space, which is iteratively called until the target domain is reached. A specific training schedule is required to alleviate the exponentiation effect of the iterations. At test time, it offers several advantages: the number of weight parameters is limited and the compositional design allows one to modulate the strength of the transformation with the number of iterations. This is useful, for instance, when the type or amount of noise to suppress is not known in advance. Experimentally, we provide proofs of concepts showing the interest of our method for many transformations. The performance of our model is comparable or better than CycleGAN with significantly fewer parameters.

Hugo Touvron, Andrea Vedaldi, Matthijs Douze et al.

This paper provides an extensive analysis of the performance of the EfficientNet image classifiers with several recent training procedures, in particular one that corrects the discrepancy between train and test images. The resulting network, called FixEfficientNet, significantly outperforms the initial architecture with the same number of parameters. For instance, our FixEfficientNet-B0 trained without additional training data achieves 79.3% top-1 accuracy on ImageNet with 5.3M parameters. This is a +0.5% absolute improvement over the Noisy student EfficientNet-B0 trained with 300M unlabeled images. An EfficientNet-L2 pre-trained with weak supervision on 300M unlabeled images and further optimized with FixRes achieves 88.5% top-1 accuracy (top-5: 98.7%), which establishes the new state of the art for ImageNet with a single crop. These improvements are thoroughly evaluated with cleaner protocols than the one usually employed for Imagenet, and particular we show that our improvement remains in the experimental setting of ImageNet-v2, that is less prone to overfitting, and with ImageNet Real Labels. In both cases we also establish the new state of the art.

Alexandre Sablayrolles, Matthijs Douze, Cordelia Schmid et al.

We want to detect whether a particular image dataset has been used to train a model. We propose a new technique, \emph{radioactive data}, that makes imperceptible changes to this dataset such that any model trained on it will bear an identifiable mark. The mark is robust to strong variations such as different architectures or optimization methods. Given a trained model, our technique detects the use of radioactive data and provides a level of confidence (p-value). Our experiments on large-scale benchmarks (Imagenet), using standard architectures (Resnet-18, VGG-16, Densenet-121) and training procedures, show that we can detect usage of radioactive data with high confidence (p<10^-4) even when only 1% of the data used to trained our model is radioactive. Our method is robust to data augmentation and the stochasticity of deep network optimization. As a result, it offers a much higher signal-to-noise ratio than data poisoning and backdoor methods.

Alexandre Sablayrolles, Matthijs Douze, Yann Ollivier et al.

Membership inference determines, given a sample and trained parameters of a machine learning model, whether the sample was part of the training set. In this paper, we derive the optimal strategy for membership inference with a few assumptions on the distribution of the parameters. We show that optimal attacks only depend on the loss function, and thus black-box attacks are as good as white-box attacks. As the optimal strategy is not tractable, we provide approximations of it leading to several inference methods, and show that existing membership inference methods are coarser approximations of this optimal strategy. Our membership attacks outperform the state of the art in various settings, ranging from a simple logistic regression to more complex architectures and datasets, such as ResNet-101 and Imagenet.

Pierre Stock, Armand Joulin, Rémi Gribonval et al.

In this paper, we address the problem of reducing the memory footprint of convolutional network architectures. We introduce a vector quantization method that aims at preserving the quality of the reconstruction of the network outputs rather than its weights. The principle of our approach is that it minimizes the loss reconstruction error for in-domain inputs. Our method only requires a set of unlabelled data at quantization time and allows for efficient inference on CPU by using byte-aligned codebooks to store the compressed weights. We validate our approach by quantizing a high performing ResNet-50 model to a memory size of 5MB (20x compression factor) while preserving a top-1 accuracy of 76.1% on ImageNet object classification and by compressing a Mask R-CNN with a 26x factor.

Guillaume Lample, Alexandre Sablayrolles, Marc'Aurelio Ranzato et al.

This paper introduces a structured memory which can be easily integrated into a neural network. The memory is very large by design and significantly increases the capacity of the architecture, by up to a billion parameters with a negligible computational overhead. Its design and access pattern is based on product keys, which enable fast and exact nearest neighbor search. The ability to increase the number of parameters while keeping the same computational budget lets the overall system strike a better trade-off between prediction accuracy and computation efficiency both at training and test time. This memory layer allows us to tackle very large scale language modeling tasks. In our experiments we consider a dataset with up to 30 billion words, and we plug our memory layer in a state-of-the-art transformer-based architecture. In particular, we found that a memory augmented model with only 12 layers outperforms a baseline transformer model with 24 layers, while being twice faster at inference time. We release our code for reproducibility purposes.

Hugo Touvron, Andrea Vedaldi, Matthijs Douze et al.

Data-augmentation is key to the training of neural networks for image classification. This paper first shows that existing augmentations induce a significant discrepancy between the typical size of the objects seen by the classifier at train and test time. We experimentally validate that, for a target test resolution, using a lower train resolution offers better classification at test time. We then propose a simple yet effective and efficient strategy to optimize the classifier performance when the train and test resolutions differ. It involves only a computationally cheap fine-tuning of the network at the test resolution. This enables training strong classifiers using small training images. For instance, we obtain 77.1% top-1 accuracy on ImageNet with a ResNet-50 trained on 128x128 images, and 79.8% with one trained on 224x224 image. In addition, if we use extra training data we get 82.5% with the ResNet-50 train with 224x224 images. Conversely, when training a ResNeXt-101 32x48d pre-trained in weakly-supervised fashion on 940 million public images at resolution 224x224 and further optimizing for test resolution 320x320, we obtain a test top-1 accuracy of 86.4% (top-5: 98.0%) (single-crop). To the best of our knowledge this is the highest ImageNet single-crop, top-1 and top-5 accuracy to date.

I. Zeki Yalniz, Hervé Jégou, Kan Chen et al.

This paper presents a study of semi-supervised learning with large convolutional networks. We propose a pipeline, based on a teacher/student paradigm, that leverages a large collection of unlabelled images (up to 1 billion). Our main goal is to improve the performance for a given target architecture, like ResNet-50 or ResNext. We provide an extensive analysis of the success factors of our approach, which leads us to formulate some recommendations to produce high-accuracy models for image classification with semi-supervised learning. As a result, our approach brings important gains to standard architectures for image, video and fine-grained classification. For instance, by leveraging one billion unlabelled images, our learned vanilla ResNet-50 achieves 81.2% top-1 accuracy on the ImageNet benchmark.

Pierre Stock, Benjamin Graham, Rémi Gribonval et al.

Modern neural networks are over-parametrized. In particular, each rectified linear hidden unit can be modified by a multiplicative factor by adjusting input and output weights, without changing the rest of the network. Inspired by the Sinkhorn-Knopp algorithm, we introduce a fast iterative method for minimizing the L2 norm of the weights, equivalently the weight decay regularizer. It provably converges to a unique solution. Interleaving our algorithm with SGD during training improves the test accuracy. For small batches, our approach offers an alternative to batch-and group-normalization on CIFAR-10 and ImageNet with a ResNet-18.

Maxim Berman, Hervé Jégou, Andrea Vedaldi et al.

MultiGrain is a network architecture producing compact vector representations that are suited both for image classification and particular object retrieval. It builds on a standard classification trunk. The top of the network produces an embedding containing coarse and fine-grained information, so that images can be recognized based on the object class, particular object, or if they are distorted copies. Our joint training is simple: we minimize a cross-entropy loss for classification and a ranking loss that determines if two images are identical up to data augmentation, with no need for additional labels. A key component of MultiGrain is a pooling layer that takes advantage of high-resolution images with a network trained at a lower resolution. When fed to a linear classifier, the learned embeddings provide state-of-the-art classification accuracy. For instance, we obtain 79.4% top-1 accuracy with a ResNet-50 learned on Imagenet, which is a +1.8% absolute improvement over the AutoAugment method. When compared with the cosine similarity, the same embeddings perform on par with the state-of-the-art for image retrieval at moderate resolutions.

Arun Mukundan, Giorgos Tolias, Andrei Bursuc et al.

We propose a multiple-kernel local-patch descriptor based on efficient match kernels from pixel gradients. It combines two parametrizations of gradient position and direction, each parametrization provides robustness to a different type of patch mis-registration: polar parametrization for noise in the patch dominant orientation detection, Cartesian for imprecise location of the feature point. Combined with whitening of the descriptor space, that is learned with or without supervision, the performance is significantly improved. We analyze the effect of the whitening on patch similarity and demonstrate its semantic meaning. Our unsupervised variant is the best performing descriptor constructed without the need of labeled data. Despite the simplicity of the proposed descriptor, it competes well with deep learning approaches on a number of different tasks.

Alexandre Sablayrolles, Matthijs Douze, Cordelia Schmid et al.

Convolutional neural networks memorize part of their training data, which is why strategies such as data augmentation and drop-out are employed to mitigate overfitting. This paper considers the related question of "membership inference", where the goal is to determine if an image was used during training. We consider it under three complementary angles. We show how to detect which dataset was used to train a model, and in particular whether some validation images were used at train time. We then analyze explicit memorization and extend classical random label experiments to the problem of learning a model that predicts if an image belongs to an arbitrary set. Finally, we propose a new approach to infer membership when a few of the top layers are not available or have been fine-tuned, and show that lower layers still carry information about the training samples. To support our findings, we conduct large-scale experiments on Imagenet and subsets of YFCC-100M with modern architectures such as VGG and Resnet.

Alexandre Sablayrolles, Matthijs Douze, Cordelia Schmid et al.

Discretizing multi-dimensional data distributions is a fundamental step of modern indexing methods. State-of-the-art techniques learn parameters of quantizers on training data for optimal performance, thus adapting quantizers to the data. In this work, we propose to reverse this paradigm and adapt the data to the quantizer: we train a neural net which last layer forms a fixed parameter-free quantizer, such as pre-defined points of a hyper-sphere. As a proxy objective, we design and train a neural network that favors uniformity in the spherical latent space, while preserving the neighborhood structure after the mapping. We propose a new regularizer derived from the Kozachenko--Leonenko differential entropy estimator to enforce uniformity and combine it with a locality-aware triplet loss. Experiments show that our end-to-end approach outperforms most learned quantization methods, and is competitive with the state of the art on widely adopted benchmarks. Furthermore, we show that training without the quantization step results in almost no difference in accuracy, but yields a generic catalyzer that can be applied with any subsequent quantizer.

Matthijs Douze, Alexandre Sablayrolles, Hervé Jégou

Similarity search approaches based on graph walks have recently attained outstanding speed-accuracy trade-offs, taking aside the memory requirements. In this paper, we revisit these approaches by considering, additionally, the memory constraint required to index billions of images on a single server. This leads us to propose a method based both on graph traversal and compact representations. We encode the indexed vectors using quantization and exploit the graph structure to refine the similarity estimation. In essence, our method takes the best of these two worlds: the search strategy is based on nested graphs, thereby providing high precision with a relatively small set of comparisons. At the same time it offers a significant memory compression. As a result, our approach outperforms the state of the art on operating points considering 64-128 bytes per vector, as demonstrated by our results on two billion-scale public benchmarks.

Alexis Conneau, Guillaume Lample, Marc'Aurelio Ranzato et al.

State-of-the-art methods for learning cross-lingual word embeddings have relied on bilingual dictionaries or parallel corpora. Recent studies showed that the need for parallel data supervision can be alleviated with character-level information. While these methods showed encouraging results, they are not on par with their supervised counterparts and are limited to pairs of languages sharing a common alphabet. In this work, we show that we can build a bilingual dictionary between two languages without using any parallel corpora, by aligning monolingual word embedding spaces in an unsupervised way. Without using any character information, our model even outperforms existing supervised methods on cross-lingual tasks for some language pairs. Our experiments demonstrate that our method works very well also for distant language pairs, like English-Russian or English-Chinese. We finally describe experiments on the English-Esperanto low-resource language pair, on which there only exists a limited amount of parallel data, to show the potential impact of our method in fully unsupervised machine translation. Our code, embeddings and dictionaries are publicly available.

Matthijs Douze, Hervé Jégou, Jeff Johnson

This paper aims at discovering meaningful subsets of related images from large image collections without annotations. We search groups of images related at different levels of semantic, i.e., either instances or visual classes. While k-means is usually considered as the gold standard for this task, we evaluate and show the interest of diffusion methods that have been neglected by the state of the art, such as the Markov Clustering algorithm. We report results on the ImageNet and the Paris500k instance dataset, both enlarged with images from YFCC100M. We evaluate our methods with a labelling cost that reflects how much effort a human would require to correct the generated clusters. Our analysis highlights several properties. First, when powered with an efficient GPU implementation, the cost of the discovery process is small compared to computing the image descriptors, even for collections as large as 100 million images. Second, we show that descriptions selected for instance search improve the discovery of object classes. Third, the Markov Clustering technique consistently outperforms other methods; to our knowledge it has never been considered in this large scale scenario.

Matthijs Douze, Arthur Szlam, Bharath Hariharan et al.

This paper considers the problem of inferring image labels from images when only a few annotated examples are available at training time. This setup is often referred to as low-shot learning, where a standard approach is to re-train the last few layers of a convolutional neural network learned on separate classes for which training examples are abundant. We consider a semi-supervised setting based on a large collection of images to support label propagation. This is possible by leveraging the recent advances on large-scale similarity graph construction. We show that despite its conceptual simplicity, scaling label propagation up to hundred millions of images leads to state of the art accuracy in the low-shot learning regime.

Naila Murray, Hervé Jégou, Florent Perronnin et al.

We consider the design of an image representation that embeds and aggregates a set of local descriptors into a single vector. Popular representations of this kind include the bag-of-visual-words, the Fisher vector and the VLAD. When two such image representations are compared with the dot-product, the image-to-image similarity can be interpreted as a match kernel. In match kernels, one has to deal with interference, i.e. with the fact that even if two descriptors are unrelated, their matching score may contribute to the overall similarity. We formalise this problem and propose two related solutions, both aimed at equalising the individual contributions of the local descriptors in the final representation. These methods modify the aggregation stage by including a set of per-descriptor weights. They differ by the objective function that is optimised to compute those weights. The first is a "democratisation" strategy that aims at equalising the relative importance of each descriptor in the set comparison metric. The second one involves equalising the match of a single descriptor to the aggregated vector. These concurrent methods give a substantial performance boost over the state of the art in image search with short or mid-size vectors, as demonstrated by our experiments on standard public image retrieval benchmarks.

Alexandre Sablayrolles, Matthijs Douze, Hervé Jégou et al.

Hashing produces compact representations for documents, to perform tasks like classification or retrieval based on these short codes. When hashing is supervised, the codes are trained using labels on the training data. This paper first shows that the evaluation protocols used in the literature for supervised hashing are not satisfactory: we show that a trivial solution that encodes the output of a classifier significantly outperforms existing supervised or semi-supervised methods, while using much shorter codes. We then propose two alternative protocols for supervised hashing: one based on retrieval on a disjoint set of classes, and another based on transfer learning to new classes. We provide two baseline methods for image-related tasks to assess the performance of (semi-)supervised hashing: without coding and with unsupervised codes. These baselines give a lower- and upper-bound on the performance of a supervised hashing scheme.