Samuel Joseph Gershman

A quantitative portrait of habituation in Stentor coeruleus

This paper quantifies habituation and potentiation in Stentor coeruleus by analyzing response profiles across stimulation frequencies and recovery periods, revealing frequency-sensitive potentiation and decoupled recovery and potentiation processes.

Dynamic belief representation and updating through learned attractor-like dynamics in the frontal cortex

The paper shows that OFC population activity in mice exhibits task-specific attractor-like dynamics during belief-state inference, mirroring Value-RNN dynamics.

Reliability-guided meta-control in intuitive physics

This study proposes a model combining predictive abstractions and mental simulation for intuitive physics, showing it matches human behavior and eye movement patterns.

The Reuse-and-Append Memory Principle: Application to Latent Cause Inference

This paper introduces the Reuse-and-Append Memory (RAM) principle to model memory modification, showing it can replicate fear conditioning phenomena without overwriting existing memories.

Policy regularization as a unifying theory of the striatal division of labor in learning

This paper proposes policy regularization as a computational framework to explain the functional division between dorsolateral and dorsomedial striatum in learning, showing it unifies multiple behavioral tasks.

Associative learning in the protozoan Stentor coeruleus

This study demonstrates associative learning in the protozoan Stentor coeruleus through Pavlovian conditioning experiments, showing a transient enhanced contraction response to weak stimuli when paired with strong ones.

Efficient coding makes and breaks Webers law

This study shows that Webers law arises from efficient coding principles and can be broken by altering stimulus frequency distributions.

An improved Stentor coeruleus genome and time-resolved transcriptomics link cyclic nucleotide-dependent kinase signaling to single-cell habituation

This study improves the Stentor coeruleus genome and uses time-resolved transcriptomics to identify a conserved signaling program involved in single-cell habituation.

Failure to classically condition planarian flatworms

This study attempts to replicate classical conditioning in planarians but fails to find evidence of learning, suggesting historic protocols may be unsuitable for these species.

Error-driven representation learning in the mesolimbic system

The paper shows that striatal activity changes in a way consistent with dopamine-driven representation learning during reinforcement learning tasks.

Fast efficient coding and sensory adaptation in gain-adaptive recurrent networks

This study introduces a gain-adaptive recurrent network model that explains adaptive tuning curves and reconciles adapter repulsion and prior attraction in sensory systems.

Phasic dopamine drives conditioned responding beyond its role in learning

This study investigates the role of dopamine in conditioned responding, finding that it directly modulates responses beyond its role in learning.

Experience shifts value updating from synaptic plasticity to neural dynamics

The paper shows that experience shifts value updating from synaptic plasticity to neural dynamics, revealing a trade-off between memory stability and behavioral flexibility.

Entorhinal cortex signals dimensions of past experience that can be generalised in a novel environment

The paper shows that the entorhinal cortex signals generalisable aspects of past experience during early exposure to novel environments, improving decision-making accuracy.

Reconciling time and prediction error theories of associative learning

This paper proposes a Bayesian theory for temporal causal learning that explains learning phenomena through timescale invariance and integrates aspects of timing and prediction error theories.

Storing long-lived memories via molecular error correction: a minimal mathematical model of Cricks memory switch

The paper develops a mathematical model of molecular error correction to improve memory storage in cells, finding that moderate error correction optimally balances memory lifetime and fidelity.

Distributed range adaptation in human parietal encoding of numbers

This study shows that human parietal cortex dynamically recalibrates number representations based on context, improving behavioral accuracy through distributed range adaptation.

Bayesian estimation yields anti-Weber variability

The paper examines Bayesian observer variability in a numerosity-estimation task and finds anti-Weber behavior, similar to human subjects, suggesting variability is influenced by prior distribution skewness.

Neural and behavioral signatures of policy compression in cognitive control

The paper presents an information-theoretic framework for understanding cognitive control costs in context-dependent decisions, showing that policy complexity increases with response deadlines and correlates with neural dimensionality.

Neural evidence that humans reuse strategies to solve new tasks

The paper shows humans reuse learned strategies for new tasks using a computational process involving successor features and generalised policy improvement, supported by neural activation patterns.

Devaluing memories of reward: A case for dopamine

This study identifies dopamine's role in devaluing reward memories through experimental and computational methods, showing its necessity for mediated devaluation and hedonic evaluation reduction.

Policy complexity suppresses dopamine responses

The paper shows that policy complexity suppresses dopamine responses to reward outcomes, reducing behavioral sensitivity to these outcomes.

The role of prospective contingency in the control of behavior and dopamine signals during associative learning

This study investigates how dopamine signals in the ventral striatum relate to contingency during associative learning in mice, finding that TD learning models with ITI state representations explain the results better than contingency-based models.

Scaling models of visual working memory to natural images

This paper develops models to explain memory errors in visual working memory tasks using natural images and finds that deep encoders predict difficulty better than baselines and replicate effects in artificial stimuli.

Emergence of belief-like representations through reinforcement learning

This paper shows that recurrent neural networks can estimate value directly from observations without explicitly forming beliefs, generating reward prediction errors similar to experimental observations.

Rate-distortion theory of neural coding and its implications for working memory

The paper applies rate-distortion theory to model visual working memory, showing it reproduces key regularities and makes novel predictions verified with monkey neuron recordings.

The Neural Architecture of Theory-based Reinforcement Learning

The paper uses theory-based reinforcement learning to analyze brain data and identifies neural mechanisms underlying theory-based learning in humans.

Dopamine mediates the bidirectional update of interval timing

This study shows dopamine signals modulate subjective time speed during interval timing, supporting a unified theory of reinforcement learning and timing.

The Role of State Uncertainty in the Dynamics of Dopamine

This study shows that sensory feedback causes dopamine to ramp up before reward, challenging conventional models and supporting the reward prediction error hypothesis through experimental and theoretical analysis.

Structure in motion: visual motion perception as online hierarchical inference

This paper proposes an online hierarchical Bayesian inference model to explain human motion perception and demonstrates its ability to qualitatively and quantitatively explain various motion stimuli.