About

Marcus Benna is an Assistant Professor at UC San Diego in the Department of Neurobiology within the School of Biological Sciences. He is associated with the Benna Lab of Theoretical and Computational Neuroscience. His research focuses on understanding neural computation, particularly in the context of reinforcement learning and information processing in neural circuits. The lab investigates how biological brains implement learning mechanisms and neural computations that are often modeled in artificial systems, aiming to bridge the gap between biological plausibility and computational efficiency. Benna's work contributes to the broader understanding of neural dynamics and learning processes, advancing the field of theoretical and computational neuroscience.

Research topics

• Computer Science
• Neuroscience
• Psychology
• Artificial Intelligence
• Biology
• Psychiatry
• Chemistry
• Machine Learning
• Cognitive psychology
• Mathematics
• Cognitive science
• Ecology

Selected publications

• Motor learning refines thalamic influence on motor cortex

  Nature · 2025-05-07 · 11 citations

  article
  PublisherDOI

• Emulating complex synapses using interlinked proton conductors

  Physical Review Applied · 2025-01-13

  article

  In terms of energy efficiency and computational speed, neuromorphic electronics based on nonvolatile memory devices are expected to be one of most promising hardware candidates for future artificial intelligence (AI). However, catastrophic forgetting, networks rapidly overwriting previously learned weights when learning new tasks, remains a pivotal obstacle in either digital or analog AI chips for unleashing the true power of brainlike computing. To address catastrophic forgetting in the context of online memory storage, a complex synapse model (the Benna-Fusi model) was proposed recently [M. K. Benna and S. Fusi, Nat. Neurosci. 19, 1697 (2016)], the synaptic weight and internal variables of which evolve following diffusion dynamics. In this work, by designing a proton transistor with a series of charge-diffusion-controlled storage components, we have experimentally realized the Benna-Fusi artificial complex synapse. Memory consolidation from coupled storage components is revealed by both numerical simulations and experimental observations. Different memory timescales for the complex synapse are engineered by the diffusion length of charge carriers and the capacity and number of coupled storage components. The advantages of the demonstrated complex synapse for both memory capacity and memory consolidation are revealed by neural network simulations of face-familiarity detection. Our experimental realization of the complex synapse suggests a promising approach to enhance memory capacity and to enable continual learning.

  PublisherDOI

• Parallel synapses with transmission nonlinearities enhance neuronal classification capacity

  PLoS Computational Biology · 2025-05-09

  articleOpen accessSenior author

  Cortical neurons often establish multiple synaptic contacts with the same postsynaptic neuron. To avoid functional redundancy of these parallel synapses, it is crucial that each synapse exhibits distinct computational properties. Here we model the current to the soma contributed by each synapse as a sigmoidal transmission function of its presynaptic input, with learnable parameters such as amplitude, slope, and threshold. We evaluate the classification capacity of a neuron equipped with such nonlinear parallel synapses, and show that with a small number of parallel synapses per axon, it substantially exceeds that of the Perceptron. Furthermore, the number of correctly classified data points can increase superlinearly as the number of presynaptic axons grows. When training with an unrestricted number of parallel synapses, our model neuron can effectively implement an arbitrary aggregate transmission function for each axon, constrained only by monotonicity. Nevertheless, successful learning in the model neuron often requires only a small number of parallel synapses. We also apply these parallel synapses in a feedforward neural network trained to classify MNIST images, and show that they can increase the test accuracy. This demonstrates that multiple nonlinear synapses per input axon can substantially enhance a neuron's computational power.

  PublisherDOI

• Discovering cognitive strategies with tiny recurrent neural networks

  Nature · 2025-07-02 · 28 citations

  articleOpen access

  Abstract Understanding how animals and humans learn from experience to make adaptive decisions is a fundamental goal of neuroscience and psychology. Normative modelling frameworks such as Bayesian inference 1 and reinforcement learning 2 provide valuable insights into the principles governing adaptive behaviour. However, the simplicity of these frameworks often limits their ability to capture realistic biological behaviour, leading to cycles of handcrafted adjustments that are prone to researcher subjectivity. Here we present a novel modelling approach that leverages recurrent neural networks to discover the cognitive algorithms governing biological decision-making. We show that neural networks with just one to four units often outperform classical cognitive models and match larger neural networks in predicting the choices of individual animals and humans, across six well-studied reward-learning tasks. Critically, we can interpret the trained networks using dynamical systems concepts, enabling a unified comparison of cognitive models and revealing detailed mechanisms underlying choice behaviour. Our approach also estimates the dimensionality of behaviour 3 and offers insights into algorithms learned by meta-reinforcement learning artificial intelligence agents. Overall, we present a systematic approach for discovering interpretable cognitive strategies in decision-making, offering insights into neural mechanisms and a foundation for studying healthy and dysfunctional cognition.

  PublisherOA PDFDOI

• Social Exclusion Amplifies Behavioral Responses to Physical Pain via Insular Neuromodulation

  bioRxiv (Cold Spring Harbor Laboratory) · 2025-05-13 · 1 citations

  preprintOpen access

  The "Pain Overlap Theory" (1) proposes that the experience of social pain overlaps with and amplifies the experience of physical pain by sharing parts of the same underlying processing systems (2-6). In humans, the insular cortex has been implicated in this overlap of physical and social pain, but a mechanistic link has not been made (2,4,5,7-9). To determine whether social pain can subsequently impact responses to nociceptive stimuli via convergent electrical signals (spikes) or convergent chemical signals (neuromodulators), we designed a novel Social Exclusion paradigm termed the Fear of Missing Out (FOMO) Task which facilitates a mechanistic investigation in mice. We found that socially-excluded mice display more severe responses to physical pain, disrupted valence encoding, and impaired neural representations of nociceptive stimuli. We performed a systematic biosensor panel and found that endocannabinoid and oxytocin signaling in the insular cortex have opposing responses during trials where mice were attending or not attending to the Social Exclusion events respectively, demonstrating distinct neuromodulatory substrates that underpin different states of Social Exclusion. We also found that intra-insular blockade of oxytocin signaling increased the response to physical pain following Social Exclusion. Together these findings suggest Social Exclusion effectively alters physical pain perception using neuromodulatory signaling in the insular cortex.

  PublisherOA PDFDOI

• Language Models Are Capable of Metacognitive Monitoring and Control of Their Internal Activations.

  PubMed · 2025-05-19

  preprintOpen access

  their activation patterns. By presenting models with sentence-label pairs where labels correspond to sentence-elicited internal activations along specific directions in the neural representation space, we demonstrate that LLMs can learn to report and control these activations. The performance varies with several factors: the number of example pairs provided, the semantic interpretability of the target neural direction, and the variance explained by that direction. These results reveal a "metacognitive space" with dimensionality much lower than the model's neural space, suggesting LLMs can monitor only a subset of their neural mechanisms. Our findings provide empirical evidence quantifying metacognitive capabilities in LLMs, with significant implications for AI safety.

  PublisherOA PDF

• Social Exclusion Amplifies Behavioral Responses to Physical Pain via Insular Neuromodulation

  Research Square · 2025-06-03 · 1 citations

  preprintOpen access
  PublisherOA PDFDOI

• Mixed selectivity: Cellular computations for complexity

  Neuron · 2024-05-09 · 89 citations

  reviewOpen accessCorresponding

  The property of mixed selectivity has been discussed at a computational level and offers a strategy to maximize computational power by adding versatility to the functional role of each neuron. Here, we offer a biologically grounded implementational-level mechanistic explanation for mixed selectivity in neural circuits. We define pure, linear, and nonlinear mixed selectivity and discuss how these response properties can be obtained in simple neural circuits. Neurons that respond to multiple, statistically independent variables display mixed selectivity. If their activity can be expressed as a weighted sum, then they exhibit linear mixed selectivity; otherwise, they exhibit nonlinear mixed selectivity. Neural representations based on diverse nonlinear mixed selectivity are high dimensional; hence, they confer enormous flexibility to a simple downstream readout neural circuit. However, a simple neural circuit cannot possibly encode all possible mixtures of variables simultaneously, as this would require a combinatorially large number of mixed selectivity neurons. Gating mechanisms like oscillations and neuromodulation can solve this problem by dynamically selecting which variables are mixed and transmitted to the readout.

  PublisherOA PDFDOI

• Deep Learning without Weight Symmetry

  arXiv (Cornell University) · 2024-05-31 · 1 citations

  preprintOpen accessSenior author

  Backpropagation (BP), a foundational algorithm for training artificial neural networks, predominates in contemporary deep learning. Although highly successful, it is often considered biologically implausible. A significant limitation arises from the need for precise symmetry between connections in the backward and forward pathways to backpropagate gradient signals accurately, which is not observed in biological brains. Researchers have proposed several algorithms to alleviate this symmetry constraint, such as feedback alignment and direct feedback alignment. However, their divergence from backpropagation dynamics presents challenges, particularly in deeper networks and convolutional layers. Here we introduce the Product Feedback Alignment (PFA) algorithm. Our findings demonstrate that PFA closely approximates BP and achieves comparable performance in deep convolutional networks while avoiding explicit weight symmetry. Our results offer a novel solution to the longstanding weight symmetry problem, leading to more biologically plausible learning in deep convolutional networks compared to earlier methods.

  PublisherOA PDF

• Linking In-context Learning in Transformers to Human Episodic Memory

  2024-01-01

  article1st authorCorresponding
  PublisherDOI

Frequent coauthors

• Stefano Fusi

  Columbia University

  34 shared

• Silvia Bernardi

  New York State Psychiatric Institute

  17 shared

• C. Daniel Salzman

  Columbia University

  15 shared

• Jérôme Munuera

  Institut Jean Nicod

  14 shared

• Kay M. Tye

  University of California, San Diego

  13 shared

• Felix H. Taschbach

  University of California, San Diego

  11 shared

• A.A.H. van Hoek

  Howard Hughes Medical Institute

  10 shared

• Jian Li

  University of Science and Technology of China

  10 shared

Labs

• Benna LabPI