Brain-optimized deep neural network models of human visual areas learn non-hierarchical representations

Learning to segment self-generated from externally caused optic flow through sensorimotor mismatch circuits

Efficient sensory detection requires the capacity to ignore task-irrelevant information, for example when optic flow patterns created by egomotion need to be disentangled from object perception. To investigate how this is achieved in the visual system, predictive coding with sensorimotor mismatch detection is an attractive starting point. Indeed, experimental evidence for sensorimotor mismatch signals in early visual areas exists, but it is not understood how they are integrated into cortical networks that perform input segmentation and categorization. Our model advances a biologically plausible solution by extending predictive coding models with the ability to distinguish self-generated from externally caused optic flow. We first show that a simple three neuron circuit produces experience-dependent sensorimotor mismatch responses, in agreement with calcium imaging data from mice. This microcircuit is then integrated into a neural network with two generative streams. The motor-to-visual stream consists of parallel microcircuits between motor and visual areas and learns to spatially predict optic flow resulting from self-motion. The second stream bidirectionally connects a motion-selective higher visual area (mHVA) to V1, assigning a crucial role to the abundant feedback connections to V1: the maintenance of a generative model of externally caused optic flow. In the model, area mHVA learns to segment moving objects from the background, and facilitates object categorization. Based on shared neurocomputational principles across species, the model also maps onto primate visual cortex. Our work extends Hebbian predictive coding to sensorimotor settings, in which the agent actively moves - and learns to predict the consequences of its own movements.

The Geometry of Low- and High-Level Perceptual Spaces

Low-level features are typically continuous (e.g., the gamut between two colors), but semantic information is often categorical (there is no corresponding gradient between dog and turtle) and hierarchical (animals live in land, water, or air). To determine the impact of these differences on cognitive representations, we characterized the geometry of perceptual spaces of five domains: a domain dominated by semantic information (animal names presented as words), a domain dominated by low-level features (colored textures), and three intermediate domains (animal images, lightly texturized animal images that were easy to recognize, and heavily texturized animal images that were difficult to recognize). Each domain had 37 stimuli derived from the same animal names. From 13 participants (9F), we gathered similarity judgments in each domain via an efficient psychophysical ranking paradigm. We then built geometric models of each domain for each participant, in which distances between stimuli accounted for participants' similarity judgments and intrinsic uncertainty. Remarkably, the five domains had similar global properties: each required 5-7 dimensions, and a modest amount of spherical curvature provided the best fit. However, the arrangement of the stimuli within these embeddings depended on the level of semantic information: dendrograms derived from semantic domains (word, image, and lightly texturized images) were more "tree-like" than those from feature-dominated domains (heavily texturized images and textures). Thus, the perceptual spaces of domains along this feature-dominated to semantic-dominated gradient shift to a tree-like organization when semantic information dominates, while retaining a similar global geometry.

How does the primate brain combine generative and discriminative computations in vision?

Vision is widely understood as an inference problem. However, two contrasting conceptions of the inference process have each been influential in research on biological vision as well as the engineering of machine vision. The first emphasizes bottom-up signal flow, describing vision as a largely feedforward, discriminative inference process that filters and transforms the visual information to remove irrelevant variation and represent behaviorally relevant information in a format suitable for downstream functions of cognition and behavioral control. In this conception, vision is driven by the sensory data, and perception is direct because the processing proceeds from the data to the latent variables of interest. The notion of "inference" in this conception is that of the engineering literature on neural networks, where feedforward convolutional neural networks processing images are said to perform inference. The alternative conception is that of vision as an inference process in Helmholtz's sense, where the sensory evidence is evaluated in the context of a generative model of the causal processes that give rise to it. In this conception, vision inverts a generative model through an interrogation of the sensory evidence in a process often thought to involve top-down predictions of sensory data to evaluate the likelihood of alternative hypotheses. The authors include scientists rooted in roughly equal numbers in each of the conceptions and motivated to overcome what might be a false dichotomy between them and engage the other perspective in the realm of theory and experiment. The primate brain employs an unknown algorithm that may combine the advantages of both conceptions. We explain and clarify the terminology, review the key empirical evidence, and propose an empirical research program that transcends the dichotomy and sets the stage for revealing the mysterious hybrid algorithm of primate vision.

High-performing neural network models of visual cortex benefit from high latent dimensionality

Geometric descriptions of deep neural networks (DNNs) have the potential to uncover core representational principles of computational models in neuroscience. Here we examined the geometry of DNN models of visual cortex by quantifying the latent dimensionality of their natural image representations. A popular view holds that optimal DNNs compress their representations onto low-dimensional subspaces to achieve invariance and robustness, which suggests that better models of visual cortex should have lower dimensional geometries. Surprisingly, we found a strong trend in the opposite direction-neural networks with high-dimensional image subspaces tended to have better generalization performance when predicting cortical responses to held-out stimuli in both monkey electrophysiology and human fMRI data. Moreover, we found that high dimensionality was associated with better performance when learning new categories of stimuli, suggesting that higher dimensional representations are better suited to generalize beyond their training domains. These findings suggest a general principle whereby high-dimensional geometry confers computational benefits to DNN models of visual cortex.

Brain-optimized inference improves reconstructions of fMRI brain activity

The release of large datasets and developments in AI have led to dramatic improvements in decoding methods that reconstruct seen images from human brain activity. We evaluate the prospect of further improving recent decoding methods by optimizing for consistency between reconstructions and brain activity during inference. We sample seed reconstructions from a base decoding method, then iteratively refine these reconstructions using a brain-optimized encoding model that maps images to brain activity. At each iteration, we sample a small library of images from an image distribution (a diffusion model) conditioned on a seed reconstruction from the previous iteration. We select those that best approximate the measured brain activity when passed through our encoding model, and use these images for structural guidance during the generation of the small library in the next iteration. We reduce the stochasticity of the image distribution at each iteration, and stop when a criterion on the "width" of the image distribution is met. We show that when this process is applied to recent decoding methods, it outperforms the base decoding method as measured by human raters, a variety of image feature metrics, and alignment to brain activity. These results demonstrate that reconstruction quality can be significantly improved by explicitly aligning decoding distributions to brain activity distributions, even when the seed reconstruction is output from a state-of-the-art decoding algorithm. Interestingly, the rate of refinement varies systematically across visual cortex, with earlier visual areas generally converging more slowly and preferring narrower image distributions, relative to higher-level brain areas. Brain-optimized inference thus offers a succinct and novel method for improving reconstructions and exploring the diversity of representations across visual brain areas.

Second Sight: Using brain-optimized encoding models to align image distributions with human brain activity

Two recent developments have accelerated progress in image reconstruction from human brain activity: large datasets that offer samples of brain activity in response to many thousands of natural scenes, and the open-sourcing of powerful stochastic image-generators that accept both low- and high-level guidance. Most work in this space has focused on obtaining point estimates of the target image, with the ultimate goal of approximating literal pixel-wise reconstructions of target images from the brain activity patterns they evoke. This emphasis belies the fact that there is always a family of images that are equally compatible with any evoked brain activity pattern, and the fact that many image-generators are inherently stochastic and do not by themselves offer a method for selecting the single best reconstruction from among the samples they generate. We introduce a novel reconstruction procedure (Second Sight) that iteratively refines an image distribution to explicitly maximize the alignment between the predictions of a voxel-wise encoding model and the brain activity patterns evoked by any target image. We use an ensemble of brain-optimized deep neural networks trained on the Natural Scenes Dataset (NSD) as our encoding model, and a latent diffusion model as our image generator. At each iteration, we generate a small library of images and select those that best approximate the measured brain activity when passed through our encoding model. We extract semantic and structural guidance from the selected images, used for generating the next library. We show that this process converges on a distribution of high-quality reconstructions by refining both semantic content and low-level image details across iterations. Images sampled from these converged image distributions are competitive with state-of-the-art reconstruction algorithms. Interestingly, the time-to-convergence varies systematically across visual cortex, with earlier visual areas generally taking longer and converging on narrower image distributions, relative to higher-level brain areas. Second Sight thus offers a succinct and novel method for exploring the diversity of representations across visual brain areas.

Reconstructing seen images from human brain activity via guided stochastic search

Visual reconstruction algorithms are an interpretive tool that map brain activity to pixels. Past reconstruction algorithms employed brute-force search through a massive library to select candidate images that, when passed through an encoding model, accurately predict brain activity. Here, we use conditional generative diffusion models to extend and improve this search-based strategy. We decode a semantic descriptor from human brain activity (7T fMRI) in voxels across most of visual cortex, then use a diffusion model to sample a small library of images conditioned on this descriptor. We pass each sample through an encoding model, select the images that best predict brain activity, and then use these images to seed another library. We show that this process converges on high-quality reconstructions by refining low-level image details while preserving semantic content across iterations. Interestingly, the time-to-convergence differs systematically across visual cortex, suggesting a succinct new way to measure the diversity of representations across visual brain areas.