Loke, J. L. | Search bioRxiv

bioRxiv2025

Shared texture-like representations, not global form, underlie deep neural network alignment with human visual processing

Deep neural networks (DNNs) excel at predicting neural responses across the visual hierarchy 1-5, a success widely interpreted as evidence of shared object recognition computations 6,7. Yet improving DNN object recognition accuracy does not reliably increase neural predictivity 8,9, and even untrained networks predict brain responses above chance 9-11. This disconnect suggests that object recognition may not drive DNN-brain alignment. Texture-like statistics are represented in both DNNs and mid-level visual cortex 12-17. In natural images, these statistics are carried by objects and backgrounds, shaping representations and recognition in both systems 18-21. Does DNN-brain alignment reflect shared sensitivity to object-related information or texture-like statistics? To dissociate these factors, we recorded EEG from 57 participants viewing natural scenes, texture-synthesized images preserving local statistics while disrupting global form and object-only images with backgrounds removed. If alignment reflects texture-like statistics, it should peak for texture-synthesized images. If it reflects object-related processing, alignment should be strongest for natural and object-only conditions, which preserve object information. We compared EEG responses with DNN activations via weighted representational similarity analysis 22,23. Texture-synthesized images yielded the strongest DNN-EEG alignment, peaking in early responses (<200 ms) and explaining up to [~]85% of noise-ceiling-normalized explainable variance versus [~]44% for natural and [~]55% for isolated objects. Crucially, object categories were more decodable for natural and object-only images than texture-synthesized images, yet these object-rich conditions showed weaker alignment. This dissociation reveals that DNNs capture the texture-statistical component of early visual responses while failing to explain later, object-related variance. HighlightsO_LITexture-synthesized images yield the strongest DNN-brain alignment across DNNs. C_LIO_LIObject categories decode better from object-rich than texture-synthesized EEG. C_LIO_LIDNN-brain alignment reflects shared texture statistics, not object-related processing. C_LI

neuroscience↗

bioRxiv2022

A critical test of deep convolutional neural networks ability to capture recurrent processing in the brain using visual masking

Recurrent processing is a crucial feature in human visual processing supporting perceptual grouping, figure-ground segmentation, and recognition under challenging conditions. There is a clear need to incorporate recurrent processing in deep convolutional neural networks (DCNNs) but the computations underlying recurrent processing remain unclear. In this paper, we tested a form of recurrence in deep residual networks (ResNets) to capture recurrent processing signals in the human brain. Though ResNets are feedforward networks, they approximate an excitatory additive form of recurrence. Essentially, this form of recurrence consists of repeating excitatory activations in response to a static stimulus. Here, we used ResNets of varying depths (reflecting varying levels of recurrent processing) to explain electroencephalography (EEG) activity within a visual masking paradigm. Sixty-two humans and fifty artificial agents (10 ResNet models of depths - 4, 6, 10, 18 and 34) completed an object categorization task. We show that deeper networks (ResNet-10, 18 and 34) explained more variance in brain activity compared to shallower networks (ResNet-4 and 6). Furthermore, all ResNets captured differences in brain activity between unmasked and masked trials, with differences starting at [~]98ms (from stimulus onset). These early differences indicated that EEG activity reflected pure feedforward signals only briefly (up to [~]98ms). After [~]98ms, deeper networks showed a significant increase in explained variance which peaks at [~]200ms, but only within unmasked trials, not masked trials. In summary, we provided clear evidence that excitatory additive recurrent processing in ResNets captures some of the recurrent processing in humans. Significance statementThe challenge of modeling recurrent processes is not trivial and the operationalization of recurrent processing is highly contested. In this paper, we tested the ability of deep residual networks (ResNets) to explain recurrent processes in the human brain. Though ResNets are feedforward networks, they have been shown to equate operations in recurrent neural networks. In this study, we show that deeper networks explained more variance in brain activity than shallower networks. However, all networks still performed far from the noise ceiling. Thus, we conclude that recurrent processing in ResNets captures a form of recurrent processing in humans though other types of recurrent processing (inhibition, multiplicative) that are not present in current regular deep neural networks (alexnet, cornet, resnet) are necessary for building better visual models.

neuroscience↗