Deep learning is closing the gap with humans on several object recognition benchmarks. Here we investigate this gap in the context of challenging images where objects are seen from unusual viewpoints. We find that humans excel at recognizing objects in unusual poses, in contrast with state-of-the-art pretrained networks (EfficientNet, SWAG, ViT, SWIN, BEiT, ConvNext) which are systematically brittle in this condition. Remarkably, as we limit image exposure time, human performance degrades to the level of deep networks, suggesting that additional mental processes (requiring additional time) take place when humans identify objects in unusual poses. Finally, our analysis of error patterns of humans vs. networks reveals that even time-limited humans are dissimilar to feed-forward deep networks. We conclude that more work is needed to bring computer vision systems to the level of robustness of the human visual system. Understanding the nature of the mental processes taking place during extra viewing time may be key to attain such robustness.

Deep learning is providing a wealth of new approaches to the old problem of novel view synthesis, from Neural Radiance Field (NeRF) based approaches to end-to-end style architectures. Each approach offers specific strengths but also comes with specific limitations in their applicability. This work introduces ViewFusion, a state-of-the-art end-to-end generative approach to novel view synthesis with unparalleled flexibility. ViewFusion consists in simultaneously applying a diffusion denoising step to any number of input views of a scene, then combining the noise gradients obtained for each view with an (inferred) pixel-weighting mask, ensuring that for each region of the target scene only the most informative input views are taken into account. Our approach resolves several limitations of previous approaches by (1) being trainable and generalizing across multiple scenes and object classes, (2) adaptively taking in a variable number of pose-free views at both train and test time, (3) generating plausible views even in severely undetermined conditions (thanks to its generative nature) -- all while generating views of quality on par or even better than state-of-the-art methods. Limitations include not generating a 3D embedding of the scene, resulting in a relatively slow inference speed, and our method only being tested on the relatively small dataset NMR. Code is available.

It is commonly observed that deep networks trained for classification exhibit class-selective neurons in their early and intermediate layers. Intriguingly, recent studies have shown that these class-selective neurons can be ablated without deteriorating network function. But if class-selective neurons are not necessary, why do they exist? We attempt to answer this question in a series of experiments on ResNet-50s trained on ImageNet. We first show that class-selective neurons emerge during the first few epochs of training, before receding rapidly but not completely; this suggests that class-selective neurons found in trained networks are in fact vestigial remains of early training. With single-neuron ablation experiments, we then show that class-selective neurons are important for network function in this early phase of training. We also observe that the network is close to a linear regime in this early phase; we thus speculate that class-selective neurons appear early in training as quasi-linear shortcut solutions to the classification task. Finally, in causal experiments where we regularize against class selectivity at different points in training, we show that the presence of class-selective neurons early in training is critical to the successful training of the network; in contrast, class-selective neurons can be suppressed later in training with little effect on final accuracy. It remains to be understood by which mechanism the presence of class-selective neurons in the early phase of training contributes to the successful training of networks.

Current state-of-the-art deep networks are all powered by backpropagation. In this paper, we explore alternatives to full backpropagation in the form of blockwise learning rules, leveraging the latest developments in self-supervised learning. We show that a blockwise pretraining procedure consisting of training independently the 4 main blocks of layers of a ResNet-50 with Barlow Twins' loss function at each block performs almost as well as end-to-end backpropagation on ImageNet: a linear probe trained on top of our blockwise pretrained model obtains a top-1 classification accuracy of 70.48%, One of the main components behind the success of deep learning is backpropagation. It remains an open question whether comparable recognition performance can be achieved with local learning rules. Previous attempts in the context of supervised learning and unsupervised learning have only been successful on small datasets like MNIST (Salakhutdinov and Hinton, 2009; Löwe et al., 2019; Ahmad et al., 2020; Ernoult et al., 2022; Lee et al., 2015) or large datasets but small networks like VGG-11 (Belilovsky et al., 2019) (67.6% top-1 accuracy on ImageNet). Being able to train models with local learning rules at scale is useful for a multitude of reasons. This approach was recently illustrated at scale in the domain of video prediction, using a stack of VAEs trained sequentially in a greedy fashion (Wu et al., 2021). From a neuroscientific standpoint, it is interesting to explore the viability of alternative learning rules to backpropagation, as it is debated whether the brain performs backpropagation (mostly considered implausible) (Lillicrap et al., 2020), approximations of backpropagation (Lillicrap et al., 2016), or relies instead on local learning rules (Halvagal and Zenke, 2022; Clark et al., 2021; Illing et al., 2021). Finally, local learning rules could unlock the possibility for adaptive computations, as each part of the network is trained to solve a subtask in isolation, naturally tuning different parts to solve different tasks (Yin et al., 2022; Baldock et al., 2021), offering interesting energy and speed trade-offs depending on the complexity of the input. There is evidence that the brain also uses computational paths that depends on the complexity of the task (e.g., Shepard and Metzler (1971)).

Self-supervised learning (SSL) is rapidly closing the gap with supervised methods on large computer vision benchmarks. A successful approach to SSL is to learn representations which are invariant to distortions of the input sample. However, a recurring issue with this approach is the existence of trivial constant representations. Most current methods avoid such collapsed solutions by careful implementation details. We propose an objective function that naturally avoids such collapse by measuring the cross-correlation matrix between the outputs of two identical networks fed with distorted versions of a sample, and making it as close to the identity matrix as possible. This causes the representation vectors of distorted versions of a sample to be similar, while minimizing the redundancy between the components of these vectors. The method is called Barlow Twins, owing to neuroscientist H. Barlow's redundancy-reduction principle applied to a pair of identical networks. Barlow Twins does not require large batches nor asymmetry between the network twins such as a predictor network, gradient stopping, or a moving average on the weight updates. It allows the use of very high-dimensional output vectors. Barlow Twins outperforms previous methods on ImageNet for semi-supervised classification in the low-data regime, and is on par with current state of the art for ImageNet classification with a linear classifier head, and for transfer tasks of classification and object detection.

A core challenge in Machine Learning is to learn to disentangle natural factors of variation in data (e.g. object shape vs. pose). A popular approach to disentanglement consists in learning to map each of these factors to distinct subspaces of a model's latent representation. However, this approach has shown limited empirical success to date. Here, we show that, for a broad family of transformations acting on images--encompassing simple affine transformations such as rotations and translations--this approach to disentanglement introduces topological defects (i.e. discontinuities in the encoder). Motivated by classical results from group representation theory, we study an alternative, more flexible approach to disentanglement which relies on distributed latent operators, potentially acting on the entire latent space. We theoretically and empirically demonstrate the effectiveness of this approach to disentangle affine transformations. Our work lays a theoretical foundation for the recent success of a new generation of models using distributed operators for disentanglement.