Caglar, L.R., Mediano, P., & Lin, B. (2026).

[ICLR HCAIR Workshop]

#inductive-bias #generalization #lossy-compression #asymmetry #perceptual judgment

Humans and modern vision models can achieve similar classification accuracy yet diverge in the structure of their errors. We show that directional confusions—systematic asymmetries in misclassification direction—reveal distinct inductive biases that are invisible to accuracy. Using matched human and deep vision model responses on a natural-image categorization task under 12 perturbation types, we quantify asymmetry in confusion matrices and link it to generalization geometry through a rate--distortion (RD) framework, summarized by efficiency (AUC) and shape descriptors (β, κ). We find that humans exhibit broad but weak asymmetries, whereas deep vision models show sparser, stronger directional collapses. Robustness training reduces global asymmetry but fails to recover the human-like breadth--strength profile of graded similarity. Mechanistic simulations further show that different asymmetry organizations can shift the RD frontier in opposite directions, even when matched for performance. Together, these results position directional confusions and RD geometry as compact, interpretable signatures of inductive bias under distribution shift.

Caglar, L.R., Mediano, P., & Lin, B. (2026).

[preprint]

#information-theory #generalization #lossy-compression

Generalization to novel visual conditions remains a central challenge for both human and machine vision, yet standard robustness metrics offer limited insight into how systems trade accuracy for robustness. We introduce a rate–distortion–theoretic framework that treats stimulus–response behavior as an effective communication channel, derives rate–distortion frontiers from confusion matrices via Blahut–Arimoto updates, and summarizes each system with two interpretable geometric signatures: slope (β) and curvature (κ), which capture the marginal cost and abruptness of accuracy–robustness trade-offs. Applying this framework to human psychophysics and 18 deep vision models under controlled image perturbations, we compare generalization geometry along two axes of model architecture and training regime (standard vs robustness-oriented). We find that both biological and artificial systems follow a common lossy-compression principle, but occupy systematically different regions of rate–distortion space. In particular, humans exhibit smoother, more flexible trade-offs, whereas modern deep networks operate in steeper, more brittle regimes (higher |β| and typically higher κ), even at matched accuracy. Across training regimes, robustness training induces systematic but dissociable shifts in β/κ, revealing cases where improved robustness or accuracy does not translate into more human-like generalization geometry. These results demonstrate that rate–distortion geometry provides a compact, model-agnostic lens for comparing generalization behavior across systems beyond standard accuracy-based metrics.

Caglar, L.R. & Lin, B. (accepted at AAAI).

[published version coming soon]

#reliableAI #NeuroAI #transparency #explainability #sustainability #mechanistic-interpretability

Calls for responsible artificial intelligence (AI) stress fairness, transparency, accountability, robustness, privacy, and sustainability, yet scale-driven systems remain opaque and energy-intensive. We propose a neuroscience-informed, function-first program with two complementary lenses. The instrumental lens selectively imports brain-derived features - sparsity, event-driven processing, local learning, modular routing, dendritic-style subunits, and neuromorphic substrates - to build sustainable and operationally viable models that meet real-world power, latency, and monitoring constraints. The epistemic lens offers representation-level interpretability tools from neuroscience for understanding complex models without full mechanistic dissection to support legibility, robustness, and contestability. We caution against “neuro-fetishism”: biological features matter only when they demonstrably advance responsible-AI objectives, and efficiency alone risks rebound without deployment guardrails. Accordingly, we outline cognitively and neurally inspired multidimensional benchmarks that pair energy and carbon audits with representation audits and cognitive stress tests. Since responsibility requires intentional deployment, these benchmarks can assist in building user-aware interfaces, oversight pathways, event-driven edge design, and regulatory scaffolding. Together, these elements chart a concrete path toward AI that is not only capable, but also accountable and sustainable by design.

Caglar, L.R.*, Corbo, J.*, Erkat, O.B., Polack, P.-O. (2025).

[COSYNE extended abstract] [ICMNS abstract] [poster]

#manifold-capacity #representational-geometry #dimensionality #learning #visual-cortex #orientation-discrimination #mice #two-photon-calcium-imaging

Disentangling sensory representations is thought to be a core computational principle of perception. In perceptual learning, geometric changes that improve the representations' linear separability have been documented in downstream and associative layes, in both artificial and biological networks. However, whether such reorganization occurs as early as the primary sensory cortices has not been investigated. Here we ask whether training mice on an orientation discrimination task reshapes the population geometry of V1 responses. Using calcium imaging in trained and naïve mice passively viewing identical stimuli, we characterized several geometric properties of population responses to task-relevant orientations — including manifold radius, centroid distance, and dimensionality — alongside measures of linear separability. We find that learning alters V1 population geometry in a manner that improves the separability of task-relevant representations, and that this reshaping is dynamic over the course of individual trials. Strikingly, whereas naïve animals show a geometry that varies linearly with stimulus angle, trained animals maintain a stable geometric structure across Go and NoGo orientations, consistent with categorical task demands. At the individual level, geometric properties are tied to behavioral performance. These results demonstrate that the geometry of cortical representations adapts to task demands as early as primary sensory cortex, giving new insights into the readout mechanisms downstream of V1, towards perceptual decisions.

Successfully identifying and interacting with objects requires a combination of high-level visual knowledge, understanding the object’s intended function, and how to grasp and manipulate them. How are these components combined into an object’s whole representation? We built a predictive encoding model to empirically investigate the componential nature of object-directed action representations, hypothesizing that they are componentially built in parietal cortex out of a set of elemental hand movements that facilitate grasping and manipulating of objects. We show that a generative model based only on a linearly weighted combination of object-associated hand movements can successfully predict the objects’ action-related fMRI activity in supramarginal gyrus. RSA searchlight analyses with three large CNNs (AlexNet, ResNet50, VGG16) further confirmed that only action-associated features and not visual features can account for the variance in parietal cortex. While this study offers important new insights into the componential mechanism of action representation and information representation more broadly, it also proposes a novel method to reverse engineer and isolate neuromotor signal from visually presented manipulable object representations.

What is the shape and structure of the space in which the brain represents objects and concepts? We investigated the neural structure of conceptual space using three publicly available neural datasets (fMRI and electrophysiology) that sample from a high number of visually presented concepts. We observed that in all cases the underlying representations shaped into a spherical manifold when using non-metric multidimensional scaling – a common dimensionality reduction technique used to visualize and inspect the underlying structure of high-dimensional data. To disentangle whether the observed scaling into a spherical manifold is an intrinsic property of the space or an artifact of the dimensionality reduction method, we investigated a multitude of variables that may affect the shape of the derived manifold in neural and simulated data. This revealed that the satisfaction of the triangle inequality (3rd metric axiom), the data’s inherent category structure, and its frequency spectrum are crucial in distinguishing between artifact and intrinsic property and indicate that spherical manifolds could be an intrinsic topological feature of how the brain represents conceptual knowledge.

A longstanding debate in cognitive science is based on two competing mathematical theories of similarity that make distinct predictions about the structure of mental representations and how to model the representational space they are stored in. Taking a mathematical modeling approach, we compared two such theories to examine whether a Euclidean space (e.g. fit by multidimensional scaling (MDS)), based on Shepards’ Universal Law of Generalization (1964, 1987), or ultrametric embedding space (e.g. clustering), based on Tversky’s set-theoretic contrast model (1977), is more appropriate to characterize the latent representational embedding space and whether they are congruent across minds and brains. Using the case of colors and letters, a human fMRI study showed that the metric structure resulting in a color wheel can be reproduced from V4 data. Surprisingly, however, the behaviorally ultrametric case of letters also revealed a metric structure in Extrastriate cortex (V2-V5), suggesting that the brain might favor a Euclidean metric representational space. 

A key challenge in neuroscience and cognitive science is to understand the organization and format of concepts representation. Part of solving this problem is to accurately characterize the structure of the representational space, which caries important information for how information is organized. We developed a mathematical and computational toolbox of methods to help characterize the representational structure and metricity of similarity spaces that can be based on behavioral data, neural data, and data from artificial neural networks.

In this paper we test the metric bias hypothesis of conceptual space proposed in Caglar et al. (2021). To test this hypothesis and have better control over the internal feature structure of the stimuli and how that relates to the space’s representational geometry, we applied a dense rather than sparse sampling by selecting conceptual subspaces with carefully constructed naturalistic categories consistent of an integral (correlated features) category structure (e.g. birds or furniture) and a set of separable (uncorrelated features) categories (Shepard et al., 1961). Both in behavioral and neural fMRI data, natural categories showed a better fit to the ultrametric space whereas the unnatural Boolean category showed a better metric fit. This revealed that there might be a perceptual-conceptual dichotomy for the resultant representations’ metricity and geometry, with conceptual categories favoring an ultrametric structure that can accommodate hierarchies and perceptual categories with a smoother similarity gradient favoring a metric space.  

In this paper we introduce a new method for decoding neural fMRI data. It is based on two assumptions, first that neural representation is distributed over networks of neurons embedded in voxel noise and second that the stimuli can be decoded as learned relations from sets of categorical stimuli. We illustrate these principles with two types of stimuli, color (wavelength) and letters (visual shape), both of which have early visual system response, but at the same time must be learned within a given function or category (i.e. color contrast, alphabet). Key to the decoding method is reducing the stimulus cross-correlation with a matched noise voxel sample by normalizing the stimulus voxel matrix. This unmasks a highly discriminative neural profile per stimulus. Projection of this new voxel space to a smaller set of dimensions, the relational information takes a unique geometric form revealing functional relationships between sets of stimuli as second-order isomorphisms.

A crucial part of understanding information representation is to understand the processes that give rise to it, namely information compression mechanisms that underly processes of abstraction and generalization and consequently facilitate learning. A different line of my research has explored this by investigating whether artificial neural networks and humans show the same attentional biases in category learning when the stimulus features are separable or integral and are matched with a categorization rule that calls for attention to a single feature (filtration) or multiple features (condensation; Garner, 1974). Comparing the learning dynamics of humans and shallow vs. deep ANNs, we found that both deep and shallow ANNs could reproduce the human-like attentional bias when given high dimensional stimuli, but that this was characterized by different internal representations and learning curves hyperbolic vs exponential curve respectively.

The claims that learning systems must build causal models and provide explanations of their inferences are not new, and advocate a cognitive functionalism for artificial intelligence. This view conflates the relationships between implicit and explicit knowledge representation. In this commentary on Lake et al.'s (2017) paper in Brain and Behavioral Sciences, we present recent evidence that neural networks do engage in model building, which is implicit, and cannot be dissociated from the learning process.