George A. Alvarez

Leading deep neural network encoding models predict visual cortical responses with nearly indistinguishable accuracy, raising the strong inference that these models have converged on the same underlying brain-aligned parameterization of natural image space. Here we demonstrate that this is not the case. We introduce axis-aligned feature accentuation, which converts each model's fitted encoding axis into graded stimulus perturbations that are predicted to parametrically control neural firing within and beyond the natural-image range. We generated over 27,500 controller stimuli from ten leading vision models and presented them to five macaques in closed-loop experiments targeting early, mid-, and high-level visual areas. Despite matched natural image predictivity, models diverged strongly in their ability to control neural firing using accentuated stimuli, revealing that most model encoding axes failed to capture the precise tuning of their corresponding neurons. The two adversarially trained models showed a consistent advantage, though adversarial robustness was only weakly predictive of neural control across other models. Instead, control was better predicted by the spatial frequency structure of the input gradient: the distribution of pixels influencing each encoding axis. Overall, these results establish neural control via axis-aligned feature accentuation as a causal method to assess the alignment between how neurons and models parameterize the visual world.

The human visual system transforms patterns of light into rich perceptual experiences, where what we see is a construction that goes beyond simple measurement. Lightness illusions—where identical parts of an image can appear dramatically different depending on context—provide a window into these processes. Here we leverage a deep learning framework to investigate the constructive processes that give rise to lightness illusions, introducing the core computational goal of edge-based image reconstruction. Specifically, we demonstrate that autoencoder models trained to reconstruct natural images based only on an edge-based image representation naturally recapitulate a wide range of lightness illusions, which were previously assumed to require distinct mechanisms, inference over lighting sources, and explicit three-dimensional scene representation. These results offer a simpler, unified account of diverse lightness phenomena as emerging naturally from surface filling-in mechanisms, and broadly provide a framework for understanding the computational principles that underlie our perception of the visual world. SIGNIFICANCE STATEMENT: The human visual system shows remarkably stable perception of objects under different viewing conditions, but it uses strategies that can be thwarted by clever visual illusions – for instance, the exact same object can appear as either white or black in different contexts. The most complex of these lightness illusions have long been taken as evidence that perception involves explicit inference about 3D scene geometry and lighting conditions. However, here we show that these illusions also emerge in deep learning models, trained simply to reconstruct natural images from sparse edge signals. Thus, our perception of the lightness of surfaces in our world may instead arise from a much more primitive computation — reconstructing surface appearance from edge responses.

A bstract: Recent progress in multimodal AI and ‘language-aligned’ visual representation learning has rekindled debates about the role of language in shaping the human visual system. In particular, the emergent ability of ‘language-aligned’ vision models (e.g. CLIP) – and even pure language models (e.g. BERT) – to predict image-evoked brain activity has led some to suggest that human visual cortex itself may be ‘language-aligned’ in comparable ways. But what would we make of this claim if the same procedures could model visual activity in a species without language? Here, we conducted controlled comparisons of pure-vision, pure-language, and multimodal vision-language models in their prediction of human (N=4) and rhesus macaque (N=6, 5:IT, 1:V1) ventral visual activity to the same set of 1000 captioned natural images (the ‘NSD1000’). The results revealed markedly similar patterns in model predictivity of early and late ventral visual cortex across both species. This suggests that language model predictivity of the human visual system is not necessarily due to the evolution or learning of language perse , but rather to the statistical structure of the visual world that is reflected in natural language.

Modular and distributed coding theories of category selectivity along the human ventral visual stream have long existed in tension. Here, we present a reconciling framework—contrastive coding—based on a series of analyses relating category selectivity within biological and artificial neural networks. We discover that, in models trained with contrastive self-supervised objectives over a rich natural image diet, category-selective tuning naturally emerges for faces, bodies, scenes, and words. Further, lesions of these model units lead to selective, dissociable recognition deficits, highlighting their distinct functional roles in information processing. Finally, these pre-identified units can predict neural responses in all corresponding face-, scene-, body-, and word-selective regions of human visual cortex, under a highly constrained sparse positive encoding procedure. The success of this single model indicates that brain-like functional specialization can emerge without category-specific learning pressures, as the system learns to untangle rich image content. Contrastive coding, therefore, provides a unifying account of object category emergence and representation in the human brain.

Deep neural network models provide a powerful experimental platform for exploring core mechanisms underlying human visual perception, such as perceptual grouping and contour integration — the process of linking local edge elements to arrive at a unified perceptual representation of a complete contour. Here, we demonstrate that feedforward, nonlinear convolutional neural networks (CNNs) can emulate this aspect of human vision without relying on mechanisms proposed in prior work, such as lateral connections, recurrence, or top-down feedback. We identify two key inductive biases that give rise to human-like contour integration in purely feedforward CNNs: a gradual progression of receptive field sizes with increasing layer depth, and a bias towards relatively straight (gradually curved) contours. While lateral connections, recurrence, and feedback are ubiquitous and important visual processing mechanisms, these results provide a computational existence proof that a feedforward hierarchy is sufficient to implement gestalt “good continuation” mechanisms that detect extended contours in a manner that is consistent with human perception.

Modular and distributed coding theories of category selectivity along the human ventral visual stream have long existed in tension. Here, we present a reconciling framework – contrastive coding – based on a series of analyses relating category selectivity within biological and artificial neural networks. We discover that, in models trained with contrastive self-supervised objectives over a rich natural image diet, category-selective tuning naturally emerges for faces, bodies, scenes, and words. Further, lesions of these model units lead to selective, dissociable recognition deficits, highlighting their distinct functional roles in information processing. Finally, these pre-identified units can predict neural responses in all corresponding face-, scene-, body-, and word-selective regions of human visual cortex, under a highly constrained sparse-positive encoding procedure. The success of this single model indicates that brain-like functional specialization can emerge without category-specific learning pressures, as the system learns to untangle rich image content. Contrastive coding, therefore, provides a unifying account of object category emergence and representation in the human brain.

Deep convolutional neural networks are a powerful model class for a range of computer vision problems, but it is difficult to interpret the image filtering process they implement, given their sheer size. In this work, we introduce a method for extracting 'feature-preserving circuits' from deep CNNs, leveraging methods from saliency-based neural network pruning. These circuits are modular sub-functions, embedded within the network, containing only a subset of convolutional kernels relevant to a target feature. We compare the efficacy of 3 saliency-criteria for extracting these sparse circuits. Further, we show how 'sub-feature' circuits can be extracted, that preserve a feature's responses to particular images, dividing the feature into even sparser filtering processes. We also develop a tool for visualizing 'circuit diagrams', which render the entire image filtering process implemented by circuits in a parsable format.

Over the last few decades, psychologists have developed precise quantitative models of human recall performance in visual working memory (VWM) tasks. However, these models are tailored to a particular class of artificial stimulus displays and simple feature reports from participants (e.g., the color or orientation of a simple object). Our work has two aims. The first is to build models that explain people’s memory errors in continuous report tasks with natural images. Here, we use image generation algorithms to generate continuously varying response alternatives that differ from the stimulus image in natural and complex ways, in order to capture the richness of people’s stored representations. The second aim is to determine whether models that do a good job of explaining memory errors with natural images also explain errors in the more heavily studied domain of artificial displays with simple items. We find that: (i) features taken from state-of-the-art deep encoders predict trial-level difficulty in natural images better than several reasonable baselines; and (ii) the same visual encoders can reproduce set-size effects and response bias curves in the artificial stimulus domains of orientation and color. Moving forward, our approach offers a scalable way to build a more generalized understanding of VWM representations by combining recent advances in both AI and cognitive modeling.

The rapid development and open-source release of highly performant computer vision models offers new potential for examining how different inductive biases impact representation learning and emergent alignment with the high-level human ventral visual system. Here, we assess a diverse set of 224 models, curated to enable controlled comparison of different model properties, testing their brain predictivity using large-scale functional magnetic resonance imaging data. We find that models with qualitatively different architectures (e.g. CNNs versus Transformers) and markedly different task objectives (e.g. purely visual contrastive learning versus vision-language alignment) achieve near equivalent degrees of brain predictivity, when other factors are held constant. Instead, variation across model visual training diets yields the largest, most consistent effect on emergent brain predictivity. Overarching model properties commonly suspected to increase brain predictivity (e.g. greater effective dimensionality; learnable parameter count) were not robust indicators across this more extensive survey. We highlight that standard model-to-brain linear re-weighting methods may be too flexible, as most performant models have very similar brain-predictivity scores, despite significant variation in their underlying representations. Broadly, our findings point to the importance of visual diet, challenge common assumptions about the methods used to link models to brains, and more concretely outline future directions for leveraging the full diversity of existing open-source models as tools to probe the common computational principles underlying biological and artificial visual systems.

Anterior regions of the ventral visual stream encode substantial information about object categories. Are top-down category-level forces critical for arriving at this representation, or can this representation be formed purely through domain-general learning of natural image structure? Here we present a fully self-supervised model which learns to represent individual images, rather than categories, such that views of the same image are embedded nearby in a low-dimensional feature space, distinctly from other recently encountered views. We find that category information implicitly emerges in the local similarity structure of this feature space. Further, these models learn hierarchical features which capture the structure of brain responses across the human ventral visual stream, on par with category-supervised models. These results provide computational support for a domain-general framework guiding the formation of visual representation, where the proximate goal is not explicitly about category information, but is instead to learn unique, compressed descriptions of the visual world.