Fake it till you make it: Learning(s) from a synthetic ImageNet clone

Recent large-scale image generation models such as Stable Diffusion have exhibited an impressive ability to generate fairly realistic images starting from a very simple text prompt. Could such models render real images obsolete for training image prediction models? In this paper, we answer part of this provocative question by questioning the need for real images when training models for ImageNet classification. More precisely, provided only with the class names that have been used to build the dataset, we explore the ability of Stable Diffusion to generate synthetic clones of ImageNet and measure how useful they are for training classification models from scratch. We show that with minimal and class-agnostic prompt engineering those ImageNet clones we denote as ImageNet-SD are able to close a large part of the gap between models produced by synthetic images and models trained with real images for the several standard classification benchmarks that we consider in this study. More importantly, we show that models trained on synthetic images exhibit strong generalization properties and perform on par with models trained on real data.

Improving the Generalization of Supervised Models

We consider the problem of training a deep neural network on a given classification task, e.g., ImageNet-1K (IN1K), so that it excels at that task as well as at other (future) transfer tasks. These two seemingly contradictory properties impose a trade-off between improving the model's generalization while maintaining its performance on the original task. Models trained with self-supervised learning (SSL) tend to generalize better than their supervised counterparts for transfer learning; yet, they still lag behind supervised models on IN1K. In this paper, we propose a supervised learning setup that leverages the best of both worlds. We enrich the common supervised training framework using two key components of recent SSL models: multi-scale crops for data augmentation and the use of an expendable projector head. We replace the last layer of class weights with class prototypes computed on the fly using a memory bank. We show that these three improvements lead to a more favorable trade-off between the IN1K training task and 13 transfer tasks. Over all the explored configurations, we single out two models: t-ReX that achieves a new state of the art for transfer learning and outperforms top methods such as DINO and PAWS on IN1K, and t-ReX* that matches the highly optimized RSB-A1 model on IN1K while performing better on transfer tasks. Project page and pretrained models: https://europe.naverlabs.com/t-rex

Concept Generalization in Visual Representation Learning

Measuring concept generalization, i.e., the extent to which models trained on a set of (seen) visual concepts can be used to recognize a new set of (unseen) concepts, is a popular way of evaluating visual representations, especially when they are learned with self-supervised learning. Nonetheless, the choice of which unseen concepts to use is usually made arbitrarily, and independently from the seen concepts used to train representations, thus ignoring any semantic relationships between the two. In this paper, we argue that semantic relationships between seen and unseen concepts affect generalization performance and propose ImageNet-CoG, a novel benchmark on the ImageNet dataset that enables measuring concept generalization in a principled way. Our benchmark leverages expert knowledge that comes from WordNet in order to define a sequence of unseen ImageNet concept sets that are semantically more and more distant from the ImageNet-1K subset, a ubiquitous training set. This allows us to benchmark visual representations learned on ImageNet-1K out-of-the box: we analyse a number of such models from supervised, semi-supervised and self-supervised approaches under the prism of concept generalization, and show how our benchmark is able to uncover a number of interesting insights. We will provide resources for the benchmark at https://europe.naverlabs.com/cog-benchmark.

Hard Negative Mixing for Contrastive Learning

Contrastive learning has become a key component of self-supervised learning approaches for computer vision. By learning to embed two augmented versions of the same image close to each other and to push the embeddings of different images apart, one can train highly transferable visual representations. As revealed by recent studies, heavy data augmentation and large sets of negatives are both crucial in learning such representations. At the same time, data mixing strategies either at the image or the feature level improve both supervised and semi-supervised learning by synthesizing novel examples, forcing networks to learn more robust features. In this paper, we argue that an important aspect of contrastive learning, i.e., the effect of hard negatives, has so far been neglected. To get more meaningful negative samples, current top contrastive self-supervised learning approaches either substantially increase the batch sizes, or keep very large memory banks; increasing the memory size, however, leads to diminishing returns in terms of performance. We therefore start by delving deeper into a top-performing framework and show evidence that harder negatives are needed to facilitate better and faster learning. Based on these observations, and motivated by the success of data mixing, we propose hard negative mixing strategies at the feature level, that can be computed on-the-fly with a minimal computational overhead. We exhaustively ablate our approach on linear classification, object detection and instance segmentation and show that employing our hard negative mixing procedure improves the quality of visual representations learned by a state-of-the-art self-supervised learning method.

Learning Visual Representations with Caption Annotations

Pretraining general-purpose visual features has become a crucial part of tackling many computer vision tasks. While one can learn such features on the extensively-annotated ImageNet dataset, recent approaches have looked at ways to allow for noisy, fewer, or even no annotations to perform such pretraining. Starting from the observation that captioned images are easily crawlable, we argue that this overlooked source of information can be exploited to supervise the training of visual representations. To do so, motivated by the recent progresses in language models, we introduce {\em image-conditioned masked language modeling} (ICMLM) -- a proxy task to learn visual representations over image-caption pairs. ICMLM consists in predicting masked words in captions by relying on visual cues. To tackle this task, we propose hybrid models, with dedicated visual and textual encoders, and we show that the visual representations learned as a by-product of solving this task transfer well to a variety of target tasks. Our experiments confirm that image captions can be leveraged to inject global and localized semantic information into visual representations. Project website: https://europe.naverlabs.com/icmlm.