Self-supervised learning (SSL) is rapidly closing the gap with supervised methods on large computer vision benchmarks. A successful approach to SSL learn embeddings which are invariant distortions of input sample. However, a recurring issue this existence trivial constant solutions. Most current avoid such solutions by careful implementation details. We propose an objective function that naturally avoids collapse measuring cross-correlation matrix between outputs two identical networks fed...

In recent years, multielectrode arrays and large silicon probes have been developed to record simultaneously between hundreds thousands of electrodes packed with a high density. However, they require novel methods extract the spiking activity ensembles neurons. Here, we new toolbox sort spikes from these large-scale extracellular data. To validate our method, performed simultaneous loose patch recordings in rodents obtain 'ground truth' data, where solution this sorting problem is known for...

Recent experimental results based on multielectrode and imaging techniques have reinvigorated the idea that large neural networks operate near a critical point, between order disorder. However, evidence for criticality has relied definition of arbitrary parameters, or models do not address dynamical nature network activity. Here we introduce novel approach to assess overcomes these limitations, while encompassing generalizing previous criteria. We find simple model describe global activity...

A bstract The vertebrate visual system is hierarchically organized to process information in successive stages. Neural representations vary drastically across the first stages of processing: at output retina, ganglion cell receptive fields (RFs) exhibit a clear antagonistic center-surround structure, whereas primary cortex (V1), typical RFs are sharply tuned precise orientation. There currently no unified theory explaining these differences layers. Here, using deep convolutional neural...

In the early visual system, cells of same type perform computation in different places field. How these code together a complex scene is unclear. A common assumption that single-type extract single-stimulus feature to form map, but this has rarely been observed directly. Using large-scale recordings rat retina, we show homogeneous population fast OFF ganglion simultaneously encodes two radically features scene. Cells close moving object quasilinearly for its position, while distant remain...

Abstract Understanding how assemblies of neurons encode information requires recording large populations cells in the brain. In recent years, multi-electrode arrays and silicon probes have been developed to record simultaneously from hundreds or thousands electrodes packed with a high density. However, these new devices challenge classical way do spike sorting. Here we method solve issues, based on highly automated algorithm extract spikes extracellular data, show that this reached near...

In many cases of inherited retinal degenerations, ganglion cells are spared despite photoreceptor cell death, making it possible to stimulate them restore visual function. Several studies have shown that is express an optogenetic protein in and make light sensitive, a promising strategy vision. However the spatial resolution optogenetically-reactivated retinas has rarely been measured, especially primate. Since also expressed axons, unclear if these neurons will only be sensitive stimulation...

Abstract One of the most striking aspects early visual processing in retina is immediate parcellation information into multiple parallel pathways, formed by different retinal ganglion cell types each tiling entire field. Existing theories efficient coding have been unable to account for functional advantages such cell-type diversity encoding natural scenes. Here we go beyond previous analyze how a simple linear model with convolutional efficiently encodes naturalistic spatiotemporal movies...

Deep networks should be robust to rare events if they are successfully deployed in high-stakes real-world applications. Here we study the capability of deep recognize objects unusual poses. We create a synthetic dataset images orientations, and evaluate robustness collection 38 recent competitive for image classification. show that classifying these is still challenge all tested, with an average accuracy drop 29.5% compared when presented upright. This brittleness largely unaffected by...

Correlations in sensory neural networks have both extrinsic and intrinsic origins. Extrinsic or stimulus correlations arise from shared inputs to the network and, thus, depend strongly on ensemble. Intrinsic noise reflect biophysical mechanisms of interactions between neurons, which are expected be robust changes Despite importance this distinction for understanding how encode information collectively, no method exists reliably separate activity data, limiting our ability build predictive...

Neural noise sets a limit to information transmission in sensory systems. In several areas, the spiking response (to repeated stimulus) has shown higher degree of regularity than predicted by Poisson process. However, simple model explain this low variability is still lacking. Here we introduce new model, with correction statistics, that can accurately predict neural spike trains stimulus. The only two parameters but reproduce observed retinal recordings various conditions. We show...

A major goal in neuroscience is to understand how populations of neurons code for stimuli or actions. While the number that can be recorded simultaneously increasing at a fast pace, most cases these recordings cannot access complete population: some carry relevant information remain unrecorded. In particular, it hard record all same type given area. Recent progress have made possible profile each neuron area thanks genetic and physiological tools, pool together from across different...

Deep learning is closing the gap with human vision on several object recognition benchmarks. Here we investigate this for challenging images where objects are seen in unusual poses. We find that humans excel at recognizing such In contrast, state-of-the-art deep networks (EfficientNet, SWAG, ViT, SWIN, BEiT, ConvNext) and large vision-language models (Claude 3.5, Gemini 1.5, GPT-4) systematically brittle poses, exception of showing excellent robustness condition. As limit image exposure...

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 latest developments self-supervised learning. We show that a pretraining procedure consisting training independently 4 main blocks layers ResNet-50 with Barlow Twins' loss function at each block performs almost as well end-to-end on ImageNet: linear probe trained top our pretrained model obtains top-1...

Coordinated activity across networks of neurons is a hallmark both resting and active behavioral states in many species, including worms, flies, fish, mice humans 1–5 . These global patterns alter energy metabolism the brain over seconds to hours, making oxygen consumption glucose uptake widely used proxies neural 6,7 However, whether changes are causally related metabolic flux intact circuits on sub-second timescales associated with behavior, unknown. Moreover, it unclear transitions...

Correlations in sensory neural networks have both extrinsic and intrinsic origins. Extrinsic or stimulus correlations arise from shared inputs to the network, thus depend strongly on ensemble. Intrinsic noise reflect biophysical mechanisms of interactions between neurons, which are expected be robust changes Despite importance this distinction for understanding how encode information collectively, no method exists reliably separate activity data, limiting our ability build predictive models...