Theory: Artificial neural networks

Talk Session: Monday, May 19, 2025, 8:15 – 10:00 am, Talk Room 2

Schedule of Events | Search Abstracts | Symposia | Talk Sessions | Poster Sessions

Talk 1, 8:15 am

Model manifold analysis suggests human ventral vision is less like an optimal classifier and more like a feature bank

Colin Conwell¹ (), Michael F. Bonner¹; ¹Johns Hopkins University

What do deep neural network models of biological vision tell us about the computational structure of the brain? A now common finding in visual computational neuroscience is that many different kinds of deep neural network models -- each with different architectures, tasks, and training diets -- are all comparably good predictors of image-evoked brain activity in the ventral visual cortex. This relative parity of seemingly diverse models may at first seem to undermine the common intuition that we can use these models to infer the key computational principles that govern the visual brain. In this work, we show to the contrary that the relative parity of models along certain dimensions does not preclude the differentiation of those same models on other dimensions -- and in particular, that metrics of manifold geometry may in certain cases reliably predict whether one model will yield better brain-alignment than another. To do this, we assess 21 manifold geometry metrics computed across a diverse set of over 1200 deep neural network models, curated to include multiple tasks, architectures, and input diets. We then use these metrics to predict how well each model aligns with occipitotemporal cortex (OTC) activity from the human fMRI Natural Scenes Dataset. We find that manifold signal-to-noise ratio (a metric previously associated with few-shot learning) is a robust predictor of downstream brain-alignment and supersedes the predictive power both of other manifold geometry metrics (i.e. manifold capacity, effective dimensionality) and of downstream task-performance (e.g. top-k recognition accuracy) across multiple different image sets (e.g. ImageNet21K versus Places365) and model comparison probes (e.g. category-supervised versus self-supervised models). These results add to a growing body of evidence that the ventral visual stream may serve more as a basis set (or feature vocabulary) for object recognition rather than as the actual locus of recognition per se.

Talk 2, 8:30 am

Human Visual Robustness Emerges from Manifold Disentanglement in the Ventral Visual Stream

Zhenan Shao^1,2,3, Yiqing Zhou⁴, Diane M. Beck^1,3; ¹Department of Psychology, University of Illinois, Urbana-Champaign, ²Department of Computer Science, University of Illinois, Urbana-Champaign, ³Beckman Institute, University of Illinois, Urbana-Champaign, ⁴Department of Physics, Cornell University

Humans effortlessly navigate the dynamic visual world, yet deep neural networks (DNNs), despite excelling in visual tasks, are surprisingly vulnerable to image perturbations that are innocuous to humans. Aligning DNN representations with human neural representations, particularly those from higher-order regions of the ventral visual stream (VVS), has been shown to improve their robustness (Shao et al., 2024). Such observation suggests that the representational space in the VVS has desirable properties that support human robustness but are absent in DNNs. One particular framework posits that the VVS achieves robust inference by progressively disentangling neural category manifolds (Dicarlo & Cox, 2007). Specifically, neural population responses to identity-preserving changes of objects form continuous manifolds in the neural state space. These manifolds are initially tangled, i.e., linearly inseparable, but become progressively disentangled across stages of the VVS, naturally resulting in robust inference. Despite its theoretical appeal, empirical evidence for this framework remains limited. Here, using a computational characterization of neural manifold statistics (Chung et al., 2018) and a 7T fMRI dataset (Allen et al., 2022), we first demonstrate that category manifolds at different stages of the human VVS show increasingly desirable geometric properties: smaller radius and compressed dimensionality, together leading to improved overall linear separability. Importantly, we show that these properties are inheritable by DNNs through neural representation alignment and indeed predict subsequent robustness gains observed in previous work. Finally, to more directly test this framework, we propose “manifold guidance”, a method that aligns DNNs to human VVS on the granularity of category manifolds, without imposing strict individual representation matching commonly adopted in previous neural alignment studies. We show that manifold guidance is capable of leading to robustness improvements in DNNs. Our findings, thus, provide compelling evidence that human visual robustness arises from the disentanglement of category manifolds in the VVS.

This work used NCSA Delta GPU through allocation SOC230011 from the Advanced Cyberinfrastructure Coordination Ecosystem: Services & Support (ACCESS) program, which is supported by National Science Foundation grants #2138259, #2138286, #2138307, #2137603, and #2138296.

Talk 3, 8:45 am

Framed RSA: honoring representational geometry and regional-mean response preferences

JohnMark Taylor¹ (), Nikolaus Kriegeskorte¹; ¹Columbia University

Representational similarity analysis (RSA) characterizes the geometry of neural activity patterns elicited by different stimuli while discarding information about neural response preferences, regional-mean activity and the absolute location or orientation of the patterns in the multivariate response space. Regional-mean activation analysis serves the complementary purpose of comparing the average population response to different stimuli. Invariance to certain aspects of the neural code is useful for comparing systems that might use superficially different encodings to implement the same computations. However, quantities such as mean activation are physiologically or mechanistically important, and so we may want for our models to predict them correctly. Here we introduce a novel analysis technique, framed RSA, which honors both the geometry and the regional-mean preferences in evaluating model-predicted representations. To achieve this, we augment the set of patterns that define the geometry by two reference patterns: the zero-point (origin) and a uniform constant pattern in the multivariate response space, enabling RSA to incorporate information about the global location, orientation, and mean activation of neural population codes. First we present the mathematical and methodological underpinnings of framed RSA, including how it interacts with different RSA analysis choices, such as the use of cross-validated dissimilarity estimates or whitened RDM comparators. Second, we show empirically that framed RSA improves accuracy for both brain region identification (using fMRI data from the Natural Scenes Dataset) and deep neural network layer identification relative to existing RSA approaches. Framed RSA thus offers the theoretical virtue of combining the strengths of two complementary and traditionally separate analysis approaches and the practical value of improved power for model-comparative inference.

This work was supported by the National Eye Institute of the NIH under award number 1F32EY033654.

Talk 4, 9:00 am

Beyond One-Way Mapping: Linking Model-Brain Asymmetry to Behavioral Predictions in Visual Object Recognition

Sabine Muzellec^1,2,3, Kohitij Kar³; ¹Brown University, ²University of Toulouse, ³York University

Advancements in artificial neural networks (ANNs) have yielded object recognition models that closely mimic the primate ventral visual pathway. Traditional evaluation metrics focus mainly on how well ANN units predict neural activity, often overlooking the bidirectional nature of this relationship. In this study, we investigate the symmetry of predictive relationships between ANN components and inferior temporal (IT) neurons and explore its implications for aligning computational models with primate behavior. We conducted large-scale neural recordings from 288 sites across the IT cortex in two macaques engaged in 45 binary object discrimination tasks using 1,320 naturalistic images. Human behavioral data were collected from 80 participants, achieving an image-level reliability of 0.89. Our analysis revealed significant asymmetries in the bidirectional predictive relationships between ANN units and neural responses. By employing linear regression and centered kernel alignment (CKA), we tagged two classes of ANN units: "best" units (top 10th percentile explained variance, EV) demonstrated significantly higher CKA values compared to all units (p<0.0001) while "worst" units (bottom 10th percentile EV) showed significantly lower CKA values. This asymmetry was consistent across multiple architectures, including Vision Transformer (ViT), ResNet50v2, Inception-v3, and AlexNet. Crucially, we found that the "best" ANN units more accurately predicted both human and macaque object discrimination performance compared to the "worst" units (p<0.05, permutation test). This relationship also remained consistent across different object categories. Interestingly, monkey IT neurons identified as "best" units as predicted by other monkey ITs also demonstrated a similar enhanced prediction of human behavior, suggesting potential shared neural mechanisms across species. Taken together, our findings underscore that developing human-like object recognition in ANNs requires optimizing neural prediction accuracy and jointly ensuring representational symmetry with biological systems.

KK was supported by funds from the Canada Research Chair Program, the Canada First Research Excellence Funds (VISTA Program), Google Research Award, and a NSERC-DG. SM was supported by CFREF (VISTA Program).

Talk 5, 9:15 am

TopoNets: Artificial Neural Network models with brain-like topography

Mayukh Deb^1,2 (), Mainak Deb, Haider Al-Tahan^1,2, N. Apurva Ratan Murty^1,2; ¹Cognition and Brain Science, School of Psychology, Georgia Tech, Atlanta, GA 30332, ²Center of Excellence in Computational Cognition, Georgia Tech, Atlanta, GA 30332

The human visual system exhibits topographic organization across multiple scales: micro-scale orientation selective pinwheels, macro-scale faces, bodies, and scene-selective regions, and large-scale biases for real-world size and animacy. Artificial neural networks (ANNs) on the other hand, lack this inherent organization. To address this problem, we introduce TopoLoss, a novel loss function for ANNs inspired by synaptic connectivity pruning mechanisms that sculpt topographic representations in the brain. The resulting models, TopoNets, exhibit brain-like topography while maintaining task performance (on Imagenet). We validated TopoLoss on convolutional and transformer-based ANN architectures (ResNet-18s, ResNet-50s, ViT-b-32), collectively TopoNets. We show that TopoNets outperform all previous topographic models on ImageNet performance while delivering similar (or higher) levels of induced topography. But what are the representational consequences of brain-like topography? We find that inducing topography drives visual representations to be lower dimensional and in-turn more brain-like. This representational change manifests in two key ways (1) improved ability to predict neural data (on BrainScore for example), and (2) in recapitulating key topographic signatures observed in the visual system (category-selective maps for example). Next we used TopoNets to examine why the brain might adopt such a characteristic topographic design. Our findings revealed that TopoNets are exceptionally parameter efficient and exhibit robustness to lesioning (L1 pruning) and downsampling. These features indicate a functional and evolutionary advantage of cortical topography in brains. Given the ubiquity of topography in the cortex, we extended TopoLoss beyond vision to language (NanoGPT, GPT-Neo-125M) and audition models (multi-task models). Remarkably, TopoNets retained high task performance while replicating key signatures of brain-like topography observed in language and auditory cortices. Taken together, TopoLoss offers a simple and flexible way to instill brain-like topography into modern ANNs. TopoNets offer a unique combination of competitive task performance and brain-like topographic organization, heralding a new generation of brain-inspired AI.

This work was funded by the NIH Pathway to Independence Award from the NEI (R00EY032603) and a startup grant from Georgia Tech (to NARM)

Talk 6, 9:30 am

Factorized convolution models for predicting and interpreting neuronal tuning in natural images

Binxu Wang^1,2,3, Carlos Ponce^1,3; ¹Harvard University, ²Kempner Institute for the Study of Natural and Artificial Intelligence, ³Harvard Medical School

Convolutional neural networks have been extensively used to model neurons in the visual systems of primates and rodents. However, this regression is ill-posed because the number of image-response pairs are often far fewer than the feature regressors. To address this, previous approaches used unsupervised feature reduction (e.g. PCA) and penalized regression (e.g. LASSO). However, these solutions discard the spatial structure of feature, usually leading to non-smooth or non-local weight, harming interpretability. We present a supervised feature reduction method that applies tensor factorization to the covariance between image features and neuronal activations. This factorization generates paired spatial masks and feature vectors, offering clearer insights into which features matter and where they are localized. The method matches the accuracy of penalized regression approaches while providing more precise localization of neuronal receptive fields. To validate our method, we performed closed-loop experiments on neurons recorded from V1, V4, and inferotemporal (IT) cortex in two primates. Using the Evolution paradigm, neurons guided image synthesis through generative networks, and the resulting image-response pairs trained a factorized convolution model. The factor structure in these models allowed further manipulation and ablation study. We shuffled different components in these models as control models and synthesized their maximal activating images. In the same recording session, we presented back these synthesized “optimal” images and measured neuronal responses. Synthesized “optimal” images from the factor model were generally more activating than those from controls; by comparison, we identified the necessary components of the factor model for each neuron. With adversarially trained backbones, we found low-rank, localized read-out weights can synthesize the preferred images as effectively as dense weights, while simplifying the images. In this way, we transformed the dense "black-box" model of a neuron into a part-based model, that was easier to describe and manipulate, helping us understand their natural image tuning.

NIH (1DP2EY035176), NSF (CAREER 2143077), Quan Pre-Doctoral Fellowship to B.W., Kempner Post-Doctoral Fellowship to B.W.