Mo Shakiba

The brain performs visual inference with extraordinary stability and efficiency—yet it remains unknown which specific biological constraints are actually necessary to produce its characteristic population dynamics. Advances in biologically constrained recurrent neural networks—from spatially embedded architectures¹² to locality-masked connectivity³, to networks trained with Hebbian⁴ or Dalean-like⁵ learning rules—suggest that imposing the right inductive biases can make artificial networks more brain-like.

Disruptions in recurrent coordination and temporal stability are hallmarks of visual cortical dysfunctions such as amblyopia⁶ and schizophrenia⁷, motivating a mechanistic exploration of how specific biological constraints support VISp dynamics. We propose to systematically isolate and combine three key classes of biological constraint—spatial embedding, locality-masked connectivity, and Hebbian/Dalean plasticity—and construct controlled model variants ranging from minimal to full biological constraints. By performing causal ablations across these variants, we will directly test which constraints are necessary and sufficient to reproduce real VISp-like population dynamics. This will provide the first mechanistic account of how specific biological constraints shape healthy cortical dynamics—and how their loss could explain visual dysfunction in developmental or pathological conditions.

Aims & Objectives

• To construct a family of recurrent neural networks with systematically varied biological constraints, spanning spatial embedding, locality-masked connectivity, and Hebbian/Dalean plasticity, enabling causal evaluation of each constraint individually and in combination.
• To identify which specific constraints are necessary and/or sufficient to reproduce empirical VISp population dynamics, by evaluating each model variant against MICrONS⁸ recordings using established neural population metrics (e.g., temporal stability, manifold geometry, entropy, dimensionality).
• To assess how selective removal of constraints degrades visual dynamics and whether these disruptions recapitulate patterns observed in developmental or pathological visual conditions. Together, these aims will provide a mechanistic framework linking specific biological constraints to both healthy and disrupted VISp dynamics.

The recent public release of the MICrONS dataset—the first fully reconstructed mouse VISp circuit with matched large-scale population recordings—makes it possible, for the first time, to evaluate biological models directly against real cortical dynamics. In parallel, advances in biologically grounded network design—including spatially embedded recurrent architectures and local Hebbian/Dalean learning rules—now make it feasible to construct models that embody specific neural constraints rather than merely approximate function. Establishing a causal framework that links particular constraints to VISp population dynamics is especially urgent, given that disrupted recurrent stability is a hallmark of many visual disorders.

To evaluate the proposed models against real VISp dynamics, model units will be converted to rate-like activity using a smooth nonlinearity for comparison with deconvolved calcium signals. In tandem, a lightweight spiking head will generate spike trains so that analyses can also be carried out both in rate and event space, ensuring comparisons are not confounded by differences in measurement modality. Using this aligned representation, similarity across metrics can be evaluated with methods such as temporal stability, manifold geometry⁹, entropy, and dimensionality. Success will be defined as reproducing the empirical range of VISp metrics within biological variability across stimuli and repetitions.

To isolate the causal role of each biological constraint (spatial embedding, locality-masked recurrent connectivity, Hebbian/Dalean plasticity), a full factorial set of model variants will be constructed: a baseline RNN with no constraints, versions with each constraint added in isolation, all pairwise combinations, and a fully constrained model. Comparing these variants will allow necessity to be defined as the degradation of VISp-like metrics when a constraint is removed, and sufficiency as the ability of a minimal subset of constraints to reproduce the dynamics, providing insight into both healthy function and how disruptions to specific constraints may underlie disrupted cortical dynamics.

In summary, by systematically dissecting the biological constraints that give rise to cortical-like dynamics, this project bridges the gap between anatomical structure, neural computation, and pathology. The resulting models will not only clarify the principles that govern visual cortical activity but also inform the design of biologically grounded artificial networks capable of learning and adapting as efficiently as the brain.

References