cortical-geometry-wiring-rnn-inductive-bias

name	cortical-geometry-wiring-rnn-inductive-bias
description	Harnessing cortical geometry, wiring, and function as inductive biases for recurrent neural networks — biologically grounded RNNs using MICrONS connectomics data (spatial coordinates, anatomical connectivity, functional relationships) to achieve superior learning performance.
version	1.0.0
category	neuroscience
tags	["RNN","cortical-geometry","connectomics","MICrONS","inductive-bias","biological-neural-networks","spatial-constraints","functional-connectivity"]
arxiv	2606.14975
authors	["Mo Shakiba","Rana Rokni","Mohammad Mohammadi","Nima Dehghani"]
published	2026-06-12T00:00:00.000Z
activation_keywords	["cortical geometry","biological RNN","MICrONS","inductive bias","spatial embedding","functional weight initialization","neuronal constraints","connectomics"]

Harnessing Cortical Geometry, Wiring, and Function as Inductive Biases for RNNs

Overview

Biologically grounded recurrent neural networks using cortical geometry, anatomical wiring, and functional relationships from MICrONS connectomics data to achieve superior learning performance on cognitive tasks.

arXiv: 2606.14975
Authors: Mo Shakiba, Rana Rokni, Mohammad Mohammadi, Nima Dehghani
Published: June 12, 2026
Categories: cs.NE, cs.AI, cs.LG, physics.data-an, q-bio.NC

Core Innovation

Demonstrates that neuronal constraints (geometry, wiring, function) as inductive biases:

• Improve RNN learning performance across cognitive tasks
• Develop low-entropy, modular, small-world organization
• Retain strong performance with positive-only weights
• Converge toward biological organizational principles

Key Methodology

1. Data Source: MICrONS Program

• ~12,000 co-registered excitatory neurons
• Dense calcium imaging + electron microscopy reconstruction
• Mouse visual cortex multiple areas
• Same animal functional-connectomics mapping

2. Biological Constraints

• Spatial coordinates: Neuron positions in cortex
• Anatomical connectivity: EM-derived wiring
• Functional relationships: Calcium imaging-derived correlations

3. Three Constraint Types

1. Functional weight initialization: Largest performance gain
2. Spatial embedding constraints: Robust additional improvements
3. Communication-aware constraints: Distance-based message passing

Architecture Design

Biological Grounding

class CorticallyConstrainedRNN:
    def __init__(self, microns_data):
        # Extract biological priors
        self.neuron_positions = microns_data.spatial_coords  # 12k neurons
        self.anatomical_wiring = microns_data.em_connectivity
        .functional_corr = microns_data.ca_imaging_correlations
        
        
        .rnn_weights = .functional_weight_init()
        
        
        .spatial_constraint = .compute_distance_penalty()