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

PLoS Computational Biology, Journal Year: 2024, Volume and Issue: 20(1), P. e1011792 - e1011792

Published: Jan. 10, 2024

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

Language: Английский

---

A Unifying Principle for the Functional Organization of Visual Cortex

bioRxiv (Cold Spring Harbor Laboratory), Journal Year: 2023, Volume and Issue: unknown

Published: May 18, 2023

A key feature of many cortical systems is functional organization: the arrangement neurons with specific properties in characteristic spatial patterns across surface. However, principles underlying emergence and utility organization are poorly understood. Here we develop Topographic Deep Artificial Neural Network (TDANN), first unified model to accurately predict multiple areas primate visual system. We analyze factors responsible for TDANN's success find that it strikes a balance between two objectives: achieving task-general sensory representation self-supervised, maximizing smoothness responses sheet according metric scales relative surface area. In turn, representations learned by TDANN lower dimensional more brain-like than those models lack constraint. Finally, provide evidence balances performance inter-area connection length, use resulting proof-of-principle optimization prosthetic design. Our results thus offer principle understanding novel view role system particular.

Language: Английский

---

Energy Guided Diffusion for Generating Neurally Exciting Images

bioRxiv (Cold Spring Harbor Laboratory), Journal Year: 2023, Volume and Issue: unknown

Published: May 20, 2023

In recent years, most exciting inputs (MEIs) synthesized from encoding models of neuronal activity have become an established method to study tuning properties biological and artificial visual systems. However, as we move up the hierarchy, complexity computations increases. Consequently, it becomes more challenging model activity, requiring complex models. this study, introduce a new attention readout for convolutional data-driven core neurons in macaque V4 that outperforms state-of-the-art task-driven ResNet predicting responses. predictive network deeper complex, synthesizing MEIs via straightforward gradient ascent (GA) can struggle produce qualitatively good results overfit idiosyncrasies model, potentially decreasing MEI's model-to-brain transferability. To solve problem, propose diffusion-based generating Energy Guidance (EGG). We show V4, EGG generates single neuron generalize better across architectures than GA while preserving within-architectures activation 4.7x less compute time. Furthermore, diffusion be used generate other neurally images, like natural images are on par with selection highly activating or image reconstructions architectures. Finally, is simple implement, requires no retraining easily generalized provide characterizations system, such invariances. Thus provides general flexible framework coding system context images.

Language: Английский

---

How well do models of visual cortex generalize to out of distribution samples?

PLoS Computational Biology, Journal Year: 2024, Volume and Issue: 20(5), P. e1011145 - e1011145

Published: May 31, 2024

Unit activity in particular deep neural networks (DNNs) are remarkably similar to the neuronal population responses static images along primate ventral visual cortex. Linear combinations of DNN unit activities widely used build predictive models Nevertheless, prediction performance these is often investigated on stimulus sets consisting everyday objects under naturalistic settings. Recent work has revealed a generalization gap how predicting synthetically generated out-of-distribution (OOD) stimuli. Here, we recent progress improving DNNs’ object recognition generalization, as well various design choices such architecture, learning algorithm, and datasets have impacted predictivity. We came surprising conclusion that none common computer vision OOD benchmarks predictivity performance. Furthermore, found adversarially robust yield substantially higher predictivity, although degree robustness itself was not score. These results suggest behavior current alone may lead more general neurons

Language: Английский

---

Heterogeneous orientation tuning in the primary visual cortex of mice diverges from Gabor-like receptive fields in primates

Cell Reports, Journal Year: 2024, Volume and Issue: 43(8), P. 114639 - 114639

Published: Aug. 1, 2024

A key feature of neurons in the primary visual cortex (V1) primates is their orientation selectivity. Recent studies using deep neural network models showed that most exciting input (MEI) for mouse V1 exhibit complex spatial structures predict non-uniform selectivity across receptive field (RF), contrast to classical Gabor filter model. Using local patches drifting gratings, we identified heterogeneous tuning varied up 90° sub-regions RF. This heterogeneity correlated with deviations from optimal filters and was consistent cortical layers recording modalities (calcium vs. spikes). In contrast, model-synthesized MEIs macaque were predominantly like, previous studies. These findings suggest emerges earlier pathway mice than primates. may provide a faster, though less general, method extracting task-relevant information.

Language: Английский

---

The mechanism at hand

Published: Dec. 23, 2024

Chapter 1 2009) go hand in with its manual capabilities.In part, we may observe this the large proportion of cortical homunculus dedicated to hands (Catani, 2017;Penfield & Boldrey, 1937).In light pivotal role body and specifically play human cognition, present thesis aims push boundaries sensorimotor neuroscience by modeling dexterity.Specifically, a total three empirical chapters, will assembly tools (Chapter 2), creation process 3), analysis 4) an ambitious top-down model that spans regions involved dexterity.We show presented can generate interesting hypotheses about neurocomputational principles are firmly grounded functional structural validity.The following introduction motivate our approach two philosophies mind: embodied enactive cognition.These reject view mind entirely discrete entities, perspective rooted Cartesian dualism (Descartes, 1985;Skirry, 2005;Thibaut, 2018) is still popular cognitive science today (Gallagher, 2023).They also computationalism, which oppose nonphysical mind, but locates cognition nervous system, where it merely implemented, not driven physicality (Shapiro, 2007;Shapiro Spaulding, 2021).In contrast both, modern philosophy spearheaded approach, rejects any type dichotomy considers brain, rest constitute as

Language: Английский

---