https://www.cell.com/trends/cognitive-sciences/fulltext/S1364-6613(21)00005-X

Neural population control via deep image synthesis

https://www.biorxiv.org/content/10.1101/2021.01.05.425426v1.full.pdf

https://papers.nips.cc/paper/2019/hash/02ed812220b0705fabb868ddbf17ea20-Abstract.html

Do Adversarially Robust ImageNet Models Transfer Better?

https://www.sciencedirect.com/science/article/pii/S009286742031388X

https://www.biorxiv.org/content/10.1101/2020.12.15.422882v2.external-links.html

https://www.biorxiv.org/content/10.1101/2021.06.29.450334v1.abstract

https://www.biorxiv.org/content/10.1101/2021.09.15.460220v1.full#F1

https://www.biorxiv.org/content/10.1101/2020.07.09.185116v1.full.pdff

https://www.sciencedirect.com/science/article/pii/S0959438821001276

https://proceedings.neurips.cc/paper/2019/file/182bd81ea25270b7d1c2fe8353d17fe6-Paper.pdf

COMPOSITIONAL ATTENTION: DISENTANGLING SEARCH AND RETRIEVAL

Cortical response to naturalistic stimuli is largely predictable with deep neural networks

Invariant neural subspaces maintained by feedback modulation

https://elifesciences.org/articles/76096#:~:text=The%20invariance%20is%20not%20present%20on%20the%20level,invariant%20neural%20subspace%20in%20spite%20of%20contextual%20variations

https://www.biorxiv.org/content/biorxiv/early/2022/06/24/2022.06.23.497415.full.pdf

Visual Search Asymmetry: Deep Nets and Humans Share Similar Inherent Biases

https://proceedings.neurips.cc/paper/2021/hash/37f0e884fbad9667e38940169d0a3c95-Abstract.html

A model of egocentric to allocentric understanding in mammalian brains

https://arxiv.org/abs/1912.06207#:~:text=12%20Dec%202019%5D-,From%20deep%20learning%20to%20mechanistic%20understanding%20in,the%20structure%20of%20retinal%20prediction&text=Recently%2C%20deep%20feedforward%20neural%20networks,output%20map%20of%20sensory%20neurons

The functional specialization of visual cortex emerges from training parallel pathways with self-supervised predictive learning

Overparameterized neural networks implement associative memory

Building Transformers from Neurons and Astrocytes

Meta Learning Backpropagation And Improving It

Gradient-based learning drives robust representations in recurrent neural networks by balancing compression and expansion

Efficient inverse graphics in biological face processing

Adversarial robustness of computational models of visual cortex

Abstract: Task-optimized convolutional neural networks (CNNs) show striking similarities to the ventral visual stream. However, human-imperceptible image perturbations, known as adversarial perturbations, can cause a CNN to make incorrect predictions. Here we suggest three properties that possibly lead to their brittleness: population response dimensionality, spatial frequency preference, and temporally-discontinuous training inputs. Theory suggests that the tolerance of a system to these perturbations could be related to the power law exponent (i.e., decay rate) of the eigenspectrum of its set of neural responses, where power law exponents closer to and larger than one would indicate a system that is more tolerant to input perturbations. We find that the eigenspectra of model representations decay slowly relative to those observed in neurophysiology and that robust models have eigenspectra that decay slightly faster and have higher power law exponents than those of non-robust models. We therefore investigated the spatial frequency tuning of artificial neurons and found that a large proportion of them preferred high spatial frequencies and that robust models had preferred spatial frequency distributions more aligned with the spatial frequency distribution measured in macaque V1 cells. Furthermore, robust models were quantitatively better models of V1 than non-robust models. Motivated by work in visual development showing that temporally-continuous visual experience improves view-invariant object recognition, we trained models on SAYCam, a large dataset of videos collected from the perspective of infants, and evaluated their adversarial robustness. We found that models trained on SAYCam were more robust than those trained on ImageNet and that incorporating temporal information further improved robustness. Overall, although CNNs are the state-of-the-art models of ventral visual processing, their brittle nature suggests that there is room for improvement and that one could potentially take inspiration from biology to improve their robustness.

Modelling human behaviour in cognitive tasks with latent dynamical systems

Experimenting with Theoretical Motor Neuroscience

Why can GPT learn in-context? Language models secretly perform gradient descent as meta-optimizers

Abstract representations emerge naturally in neural networks trained to perform multiple tasks

Motor cortex signals for each arm are mixed across hemispheres and neurons yet partitioned within the population response

https://elifesciences.org/articles/46159#:~:text=Neuroscience-,Motor%20cortex%20signals%20for%20each%20arm%20are%20mixed%20across%20hemispheres,partitioned%20within%20the%20population%20response

Working memory control dynamics follow principles of spatial computing

Research Presentation: Modeling Visual search in Humans