Biologically grounded neocortex computational primitives implemented on neuromorphic hardware improve vision...

Area of Science:

• Computational Neuroscience
• Artificial Intelligence
• Neuroscience

Background:

• Advancing neuro-inspired AI (NeuroAI) requires understanding brain computation and implementing it in hardware and deep learning.
• Neocortical microcircuits, particularly in the mouse primary visual cortex, utilize diverse interneuron classes for complex dynamics.

Purpose of the Study:

• To model neocortical microcircuits using an experimentally constrained, biophysically realistic approach.
• To investigate the role of four interneuron classes (Parvalbumin, Somatostatin, vasoactive intestinal peptide, LAMP5) in implementing soft winner-take-all (sWTA) circuit dynamics.
• To bridge biological computation with neuromorphic hardware and deep learning architectures.

Main Methods:

• Developed a conductance-based network model of mouse visual cortex layers 2-3, grounded in in vitro physiology.
• Implemented a competitive-cooperative motif to achieve sWTA dynamics, enabling selective input amplification and suppression.
• Mapped the sWTA motif onto IBM's TrueNorth neuromorphic chip using a gain-matching strategy.
• Integrated the sWTA circuit as a preprocessing filter within a Vision Transformer architecture.

Main Results:

• The sWTA circuit demonstrated gain modulation, signal restoration, and context-dependent multistability.
• Neuromorphic implementation revealed correspondences between cell-type roles and hardware primitives.
• Sparse sWTA modules facilitated persistent up-states and working memory approximation.
• Embedding the sWTA filter in a Vision Transformer enhanced out-of-distribution generalization by up to 20% and reduced training compute.

Conclusions:

• The study provides a roadmap for integrating biophysically grounded cortical computation into NeuroAI systems.
• The sWTA circuit serves as a key motif for enhancing AI performance and efficiency.
• This work demonstrates the potential of combining neuroscience principles with AI for practical advancements.