Blake A. Richards

This study examines neural responses in mouse visual cortex to natural and artificial visual stimuli, finding increased activity in response to natural mismatches but no change in dendritic-somatic coupling.

The paper shows that distinct inputs to the CA1 region of the hippocampus support different memory resolutions by altering neural dimensionality and feedforward inhibition.

RetINaBox provides a low-cost open-source tool for hands-on learning about visual system feature detectors through interactive experiments.

This paper introduces exponentiated gradient (EG) as a learning algorithm for artificial neural networks that better aligns with biological principles than gradient descent.

This study shows that incorporating top-down feedback in neural networks mimics human brain architecture and introduces a visual bias in audiovisual tasks without impairing performance.

This paper proposes the oneirogen hypothesis, suggesting classical psychedelics induce hallucinations by shifting neural activity toward dream-like states, and validates this through simulations using a wake-sleep neural network model.

The paper uses a computational model to show that developmental stochasticity in neural wiring increases population fitness by generating phenotypic variability in unpredictable environments.

This study shows that sensory information increases in primary somatosensory cortex during learning, while choice information emerges later and relies on population coding.

This paper shows that predictive learning in neural networks can generate hippocampal-like spatial representations and replay, suggesting a unifying theory for hippocampal function and AI.

The paper shows that cortical pyramidal neurons exhibit distinct unexpected event signals in somata and distal apical dendrites that change over days, suggesting a role in hierarchical computation.

The paper identifies visual feature preferences of neurons in mouse visual cortical areas using generative AI and closed-loop recordings, revealing how these preferences are organized to encode own-stimuli more effectively.

The paper identifies two distinct neural codes, burst fraction and event rate, that convey stimulus information and error signals during learning, with burst fraction showing stronger modulation and temporal alignment with behavioral accuracy.

This paper investigates whether neurogenesis acts as a regularizer in neural networks, finding it enhances generalization in a category learning task.

This paper uses human brain connectomes to train neuromorphic networks, showing that critical dynamics and modular organization enhance cognitive task performance.

This paper proposes parallel and recurrent cascade models to capture sub-cellular neural computation and integrates them into artificial neural networks for complex tasks.

This study shows that memory demands, not task temporal structure, are necessary for time cells to encode sensory information in DRL agents.

A 3D ResNet trained to estimate an agent's own motion from simulated visual input predicts selectivity of primate dorsal-stream neurons, whereas action-recognition networks on naturalistic video do not.

Training a two-pathway deep neural network with a single self-supervised predictive loss function models both dorsal and ventral pathways of mouse visual cortex and outperforms competing models.

This study shows that burst-dependent synaptic plasticity enables coordination of learning in hierarchical circuits by solving complex tasks through simulations and mathematical analysis.

The paper introduces VaLPACa, a variational ladder autoencoder that disentangles hierarchical dynamics in calcium imaging data by separating calcium flux from neural computation.

Splitting network units into excitatory and inhibitory populations need not cost learning: initialize inhibition to centre and standardize excitatory activity, and scale inhibitory updates by Fisher information.

The paper compares linear and nonlinear model performance on brain imaging data and standard datasets, finding linear models outperform nonlinear ones in brain scans despite larger sample sizes.

The paper investigates the role of parvalbumin and somatostatin interneurons in spike-time synchronization in the neocortex, finding frequency-specific contributions to spatio-temporal synchronization.

The paper investigates how different inhibitory interneuron subtypes in the neocortex respond to information encoded in spike synchrony versus rate.

This paper develops a neural network model showing how cortical circuits learn to ignore irrelevant inputs through inhibitory plasticity, and how disruptions to this process may contribute to schizophrenia.