Reading Group: Neuroscience × Machine Learning

One of the first large-scale brain-network mapping studies showing hierarchical correspondences between DNN layers and human visual processing in both space and time.

Showed that recurrent architectures better capture time-resolved ventral-stream dynamics than purely feedforward models, supporting a central role for recurrence in vision.

A striking emergence result: a task-optimized attention model learns retinal-like eccentricity-dependent sampling without imposing it by hand.

Introduced the Generative Query Network, showing how agents can infer compact latent scene representations and render novel viewpoints.

Linked delayed recognition under occlusion and masking effects to recurrent computations; recurrent models captured behavior and neural timing better than feedforward ones.

Compared representation geometry in ViTs and CNNs, highlighting earlier global integration in ViTs and stronger layer-wise uniformity.

Showed that explicit relation modules can dramatically improve relational reasoning capabilities in otherwise strong convolutional architectures.

Combined synaptic stabilization with context-dependent sparse gating to reduce interference across tasks, offering an accessible continual-learning recipe.

Used GPT-2-derived prediction measures with neural data to show that language comprehension involves concurrent multi-level predictive processing.

Demonstrated that goal-directed agents can develop grid-like codes that support vector navigation and shortcuts, linking function to entorhinal-style representations.

Extended goal-driven modeling beyond primate vision to rodent somatosensation, broadening the NeuroAI template across species and sensory systems.

Showed that single-neuron input-output mappings can require deep temporal architectures, emphasizing the computational richness of biophysical neurons.

A methodological stress test for neuroscience analyses, showing that standard tools can recover structure yet still miss true mechanistic understanding.

Revisited hierarchical brain-network correspondence using a direct interfacing framework, showing richer cross-stage category information than strict feedforward mappings suggest.

Classic psycholinguistics paper showing how recurrent dynamics can capture grammatical regularities from sequential input.

Directly compared human and machine multi-task learning, highlighting factorized representations as a key ingredient for reduced interference.

Introduced feedback alignment, showing useful credit assignment can emerge without exact weight symmetry required by standard backpropagation.

Showed that deep image synthesis can drive targeted V4 neural populations beyond natural response levels, revealing highly specific and sometimes counterintuitive features.

Used diffusion models with CLIP-aligned representations to reconstruct viewed images from fMRI, including ROI-specific reconstructions of hierarchical feature selectivity.

Estimated the data-efficiency gap between human visual learning and current SSL pipelines, arguing that modern systems remain far more data hungry.

A landmark multi-area recording paper showing interacting bottom-up sensory and top-down task-information flows underlying flexible decisions.

Reported that robustly generalizing models are less dependent on single highly selective directions, informing both theory and regularization practice.

Linked DQN representations to fMRI activity and behavior during Atari play, connecting RL state abstractions with distributed sensorimotor brain dynamics.