We study how the organization of representations shapes learning, generalization, and the ability to perform many tasks at once, in brains and in artificial neural networks. We combine the geometry and topology of representations with models from cognitive science to explain trade-offs, such as why representations that generalize well limit how much can be processed at once. Examples include the capacity limits of cognition, the trade-off between generalization and identification, failures of vision-language models, and models that predict activity in visual cortex.
Featured · ICLR · 2026
Intelligent systems must deploy internal representations that are simultaneously structured — to support broad generalization — and selective — to preserve input identity.
Nurisso, M., Fernando, J., Deshpande, R., Perotti, A., Marjieh, R., Frankland, S. M., Lewis, R. L., Webb, T. W., Campbell, D., Vaccarino, F., Cohen, J. D., Petri, G.
Frankland, S. M., Marjieh, R., Nurisso, M., Fluegemann, J., Webb, T. W., Petri, G., Lewis, R. L., Cohen, J. D.
The striking constraints of some human cognitive processes stand in stark contrast to the near limitless capability of others. While we can acquire and flexibly use vast amounts of information, the amount we can process at any one time is often stiflingly limited: for example the number of items we can hold in working memory or the number of tasks that can be performed at once. Here, we integrate ideas from information theory, cognitive science, and neuroscience to offer a unified account of why processing is often so limited. We argue that this reflects a fundamental tradeoff between generalization—how effectively existing representations can be used in novel settings—and how many distinct representations can be processed in parallel. Representations that best promote strong forms of generalization — a characteristically human cognitive strength — come at the expense of surprisingly strict limits in the number of items that can be processed at once, an equally characteristic human weakness. We refer to this as the “curse of generalization.” We formulate this first in information-theoretic terms, and then in process models, including a neural network model of classic tasks used to demonstrate strict limits in human processing capacity. This tension offers a potential explanation for a range of phenomena — from performance on the tasks on which we focus, to representational learning and skill acquisition more broadly — as well as the performance of modern machine learning architectures that exhibit generalization capabilities comparable to humans.
Nurisso, M., Fernando, J., Deshpande, R., Perotti, A., Marjieh, R., Frankland, S. M., Lewis, R. L., Webb, T. W., Campbell, D., Vaccarino, F., Cohen, J. D., Petri, G.
Intelligent systems must deploy internal representations that are simultaneously structured — to support broad generalization — and selective — to preserve input identity. For any model whose representational similarity between inputs decays with finite semantic resolution, we derive closed-form expressions that pin its probability of correct generalization and identification to a universal Pareto front independent of input space geometry. A minimal ReLU network trained end-to-end reproduces these laws: during learning a resolution boundary self-organizes and empirical trajectories closely follow theoretical curves. The same limits persist in two markedly more complex settings — a convolutional neural network and state-of-the-art vision-language models — confirming that finite-resolution similarity is a fundamental emergent informational constraint.
Savietto, D., Campbell, D., Panisson, A., Nurisso, M., Petri, G., Cohen, J. D., Perotti, A.
We analyze the representational geometry of open-weight vision-language models to understand their failures in multi-object tasks such as hallucinating non-existent elements or failing to identify the most similar objects among distractions. By distilling concept vectors and validating them through steering interventions, we show that geometric overlap between these vectors strongly correlates with specific error patterns.
Nurisso, M., Leroy, P., Petri, G., Vaccarino, F.
Understanding the properties of the parameter space in feed-forward ReLU networks is critical for effectively analyzing and guiding training dynamics. After initialization, training under gradient flow decisively restricts the parameter space to an algebraic variety that emerges from the homogeneous nature of the ReLU activation function. In this study, we examine two key challenges associated with feed-forward ReLU networks built on general directed acyclic graph (DAG) architectures: the (dis)connectedness of the parameter space and the existence of singularities within it. We extend previous results by providing a thorough characterization of connectedness, highlighting the roles of bottleneck nodes and balance conditions associated with specific subsets of the network. Our findings clearly demonstrate that singularities are intricately connected to the topology of the underlying DAG and its induced sub-networks. We discuss the reachability of these singularities and establish a principled connection with differentiable pruning. We validate our theory with simple numerical experiments.
Brondetta, A., Bizyaeva, A., Lucas, M., Petri, G., Musslick, S.
This simulation study examines how diversity in cognitive stability and flexibility among individuals improves collaborative task switching. Findings show that heterogeneous groups excel in high-switch environments, especially when the most flexible agents receive switch instructions, suggesting benefits for cognitive heterogeneity in clinical and educational applications.
Celeghin, A., Borriero, A., Orsenigo, D., Diano, M., Méndez Guerrero, C. A., Perotti, A., Petri, G., Tamietto, M.
This paper examines Convolutional Neural Networks (CNNs) as in silico models of the primate visual system, highlighting structural and functional parallels. It discusses challenges in modeling visual processing pathways and proposes architectural constraints to align CNNs more closely with biological vision, extending their applicability beyond object recognition.
Azeglio, S., Poetto, S., Savant, L., Nurisso, M.
Computational models of vision have traditionally been developed in a bottom-up fashion, by hierarchically composing a series of straightforward operations - i.e. convolution and pooling - with the aim of emulating simple and complex cells in the visual cortex, resulting in the introduction of deep convolutional neural networks (CNNs). Nevertheless, data obtained with recent neuronal recording techniques support that the nature of the computations carried out in the ventral visual stream is not completely captured by current deep CNN models. To fill the gap between the ventral visual stream and deep models, several benchmarks have been designed and organized into the Brain-Score platform, granting a way to perform multi-layer (V1, V2, V4, IT) and behavioral comparisons between the two counterparts. In our work, we aim to shift the focus on architectures that take into account lateral recurrent connections, a ubiquitous feature of the ventral visual stream, to devise adaptive receptive fields. Through recurrent connections, the input s long-range spatial dependencies can be captured in a local multi-step fashion and, as introduced with Gated Recurrent CNNs (GRCNN), the unbounded expansion of the neuron s receptive fields can be modulated through the use of gates. In order to increase the robustness of our approach and the biological fidelity of the activations, we employ specific data augmentation techniques in line with several of the scoring benchmarks. Enforcing some form of invariance, through heuristics, was found to be beneficial for better neural predictivity.
Petri, G., Musslick, S., Dey, B., Ozcimder, K., Turner, D., Ahmed, N.K., Willke, T.L., Cohen, J.D.
The ability to learn new tasks and generalize to others is a remarkable characteristic of both human brains and recent artificial intelligence systems. The ability to perform multiple tasks simultaneously is also a key characteristic of parallel architectures, as is evident in the human brain and exploited in traditional parallel architectures. Here we show that these two characteristics reflect a fundamental tradeoff between interactive parallelism, which supports learning and generalization, and independent parallelism, which supports processing efficiency through concurrent multitasking. Although the maximum number of possible parallel tasks grows linearly with network size, under realistic scenarios their expected number grows sublinearly. Hence, even modest reliance on shared representations, which support learning and generalization, constrains the number of parallel tasks. This has profound consequences for understanding the human brain’s mix of sequential and parallel capabilities, as well as for the development of artificial intelligence systems that can optimally manage the tradeoff between learning and processing efficiency.
