An adaptive neuromorphic model of ocular dominance map using floating gate 'synapse'

Area of Science:

• Neuroscience
• Electrical Engineering
• Computer Science

Background:

• Cortical cells process information through weighted sums and adaptation.
• Competitive learning mechanisms are crucial for feature selection in biological systems.
• Analogue Very-Large-Scale Integration (VLSI) offers efficient implementation of neural models.

Purpose of the Study:

• To present a novel analogue CMOS design of a cortical cell.
• To implement competitive learning and feature selectivity using adaptation dynamics.
• To demonstrate pattern formation capabilities in a 2-D grid.

Main Methods:

• Designed an analogue CMOS cortical cell with a weighted sum input computation.
• Utilized floating gate pFET 'synapse' adaptation for time-staggered winner-take-all competitive learning.
• Incorporated a learning rate and bias for regulating adaptation and resource limitation.
• Embedded feature-selective cells in a 2-D Resistor-Capacitor (RC) grid.

Main Results:

• The cell demonstrated feature selectivity by favoring specific input patterns after learning.
• The 2-D RC grid of cells exhibited symmetry breaking pattern formation.
• The design showed adaptability and long-term memory characteristics similar to biological networks.

Conclusions:

• The novel analogue CMOS cortical cell effectively performs competitive learning and feature selection.
• The design is suitable for analogue VLSI implementation of Self-Organizing Feature Map (SOFM) models.
• The cell's behavior closely mimics biological neural networks, suggesting potential for advanced neuromorphic computing.