Probing Human Visual Robustness with Neurally-Guided Deep Neural Networks

TL;DR

The paper tackles why humans exhibit robust vision while DNNs remain vulnerable to tiny perturbations. It introduces neurally guided DNNs aligned to multiple regions along the ventral visual stream, revealing a hierarchy of robustness that strengthens with higher-order ROI guidance. By analyzing neural manifold geometry with MFTMA and introducing manifold guidance, the authors show that smaller, more linearly separable category manifolds emerge along the VVS and transfer these properties to DNNs, supporting robustness gains. The work suggests that leveraging hierarchical neural representations and manifold geometry can guide the development of more resilient AI systems, and it discusses the potential role of top-down feedback in shaping such robustness.

Abstract

Humans effortlessly navigate the dynamic visual world, yet deep neural networks (DNNs), despite excelling at many visual tasks, are surprisingly vulnerable to minor image perturbations. Past theories suggest that human visual robustness arises from a representational space that evolves along the ventral visual stream (VVS) of the brain to increasingly tolerate object transformations. To test whether robustness is supported by such progression as opposed to being confined exclusively to specialized higher-order regions, we trained DNNs to align their representations with human neural responses from consecutive VVS regions while performing visual tasks. We demonstrate a hierarchical improvement in DNN robustness: alignment to higher-order VVS regions leads to greater improvement. To investigate the mechanism behind such robustness gains, we test a prominent hypothesis that attributes human robustness to the unique geometry of neural category manifolds in the VVS. We first reveal that more desirable manifold properties, specifically, smaller extent and better linear separability, indeed emerge across the human VVS. These properties can be inherited by neurally aligned DNNs and predict their subsequent robustness gains. Furthermore, we show that supervision from neural manifolds alone, via manifold guidance, is sufficient to qualitatively reproduce the hierarchical robustness improvements. Together, these results highlight the critical role of the evolving representational space across VVS in achieving robust visual inference, in part through the formation of more linearly separable category manifolds, which may in turn be leveraged to develop more robust AI systems.

Figures (13)

• Figure 1: (A) Illustration of anatomical locations of seven ROIs from Subject-1 used for neural guidance. (B) Architecture of the neurally guided ResNet, trained with a task head for image classification and a neural head to match neural responses generated by neural predictors.
• Figure 2: (A) Left: Top-1 classification accuracy of all models under $l_\infty$-based PGD attacks of varying strength $\epsilon$. “Clean” denotes accuracy on unperturbed images. Right: Robustness improvements summarized as the difference in area under the accuracy curve ($\Delta$AUC) for each model relative to the baseline (None model), computed across all $\epsilon$ values used (see Methods \ref{['sec:methods-att']}). (B) Transfer attack accuracy vs. native robustness improvement ($\Delta$AUC) for NG- (red diamonds) and WD-models (blue circles). Dashed lines show linear fits with slopes $\theta$ annotated (see Results \ref{['sec:results-surface']}). (C) 3D MDS visualization of representational space similarity among 16 models and 7 human ROIs. Proximity indicates similarity in the representational space. Dashed stems are added for visualization of the relative positions of each circle.
• Figure 3: (A) Isomap visualization of three example category manifolds in V1, V4, and TO representational spaces (note the shrinkage in axis scales indicating smaller extent) (B) Category manifold extent ($\downarrow$) and linear separability ($\uparrow$) estimated using MFTMA chung2018classification for human brain ROIs (left), neural predictors (middle), and neurally-guided (NG) models (right). Error bars represent 95% Confidence Interval (CI) of extent across categories. Dashed lines mark the chance level for separability. (C) Manifold extent and linear separability in NG models predict their robustness improvement ($\Delta$AUC): NG models with less diffuse (smaller extent) and more separable manifolds show greater $\Delta$AUC.
• Figure 4: (A) Illustration of manifold guidance that trains DNN with a task head for image classification and a neural head to match category manifold properties with the corresponding neural manifold ones estimated using neural predictors (B) Left: Top-1 classification accuracy of models trained with manifold guidance and the baseline None model under $l_\infty$-based PGD attacks. “Clean” denotes accuracy on unperturbed images. Right: Robustness improvements summarized as $\Delta$AUC relative to the baseline None model. (C) Correlation of robustness improvement $\Delta$AUC for DNNs trained with manifold guidance vs. neural guidance. Improvement hierarchy is preserved despite the small improvement magnitude.
• Figure 5: Schematic illustration of expected robustness improvements hierarchy across the seven ROIs included in our study.