Haim Sompolinsky

Neural representational geometry of a joint code for stimulus category and category-independent features

The paper investigates how neural representations can jointly encode stimulus categories and category-independent features, showing that regression-relevant geometry can be optimized without compromising classification geometry.

A complementary learning system for continual episodic memory in large language models

This paper introduces a complementary learning system within a large language model to simulate continual episodic memory, achieving high accuracy in recall and question answering.

Symmetries and continuous attractors in disordered neural circuits

This paper shows that continuous-attractor dynamics can operate in complex mammalian neural circuits despite apparent heterogeneity in neural responses.

Associative memory of structured knowledge

This paper develops a model for storing and retrieving structured knowledge in recurrent neuronal networks using vector symbolic architectures and associative memory rules.

A Minimum Perturbation Theory of Deep Perceptual Learning

This paper develops a statistical mechanics theory of perceptual learning, showing that minimum perturbation plasticity explains synaptic changes and neural representations in deep hierarchical networks.

Stochastic Consolidation of Lifelong Memory

This paper introduces a model for continual learning in a recurrent neural network that combines Hebbian learning, synaptic decay, and a novel memory consolidation mechanism to avoid catastrophic forgetting.

The Geometry of Concept Learning

The paper proposes a geometric framework for few-shot learning using neural manifolds and a single plastic neuron, demonstrating its effectiveness on visual concepts and linking neurophysiology to behavior.

The spectrum of covariance matrices of randomly connected recurrent neuronal networks

The paper analyzes the covariance matrix spectrum of randomly connected recurrent neuronal networks to understand connectivity-dynamics relationships and provides theoretical predictions for comparison with experimental data.

Object manifold geometry across the mouse cortical visual hierarchy

Recordings of thousands of mouse visual cortex neurons reveal that lateral areas LM, LI, and AL carry higher linearly decodable object information associated with reduced object manifold dimension and radius.

Separability and Geometry of Object Manifolds in Deep Neural Networks

This study examines how classification capacity in deep neural networks improves with hierarchy, showing that changes in object manifold geometry underlie this improvement.

Functional diversity among sensory neurons from efficient coding principles

The paper investigates how functional diversity in sensory neurons improves information processing by applying efficient coding principles to model optimal population coding of sensory variables.

Neural correlates of learning pure tones versus natural sounds in the auditory cortex

The paper investigates how perceptual learning of pure tones versus natural sounds affects neural plasticity in the auditory cortex, finding that natural sounds induce sparsening and decorrelation, improving discrimination of complex sounds.

Coherent chaos in a recurrent neural network with structured connectivity

This paper introduces a model of recurrent neural networks with structured connectivity that generates spatially correlated chaotic activity, showing how structured connections can produce coherent fluctuations across the network.

Brainwide organization of neuronal activity and convergent sensorimotor transformations in larval zebrafish

The paper analyzes whole-brain neuronal activity in larval zebrafish during visual stimuli and motor output, revealing functional brain nuclei and convergent sensorimotor transformations.