Self-similar Neural Networks Based on a Kohonen Learning Rule

Area of Science:

• Computational neuroscience
• Artificial intelligence
• Visual system development

Background:

• Mammalian visual systems display remarkable structural regularity, notably in the visual cortex, as shown by Hubel and Wiesel.
• This regularity is hypothesized to arise from structured changes in visual input caused by environmental motion.
• Understanding the developmental principles of visual system structure is crucial for neuroscience and AI.

Purpose of the Study:

• To investigate how structured visual input, specifically related to motion, can lead to the evolution of regular structures in artificial neural networks.
• To explore the role of motion-prediction error in shaping network architecture.
• To determine if simple learning rules can generate complex visual system-like structures.

Main Methods:

• Simulating artificial neural networks driven by transformations reflecting environmental motion, not random input.
• Analyzing motion-prediction error as a key factor in network learning.
• Testing network evolution in both one and two-dimensional transformation spaces.

Main Results:

• Networks evolved significant structural regularity that mirrored the symmetry groups of the applied transformations.
• In two-dimensional networks, specific transformation groups led to the development of foveal-like structures.
• Preliminary results suggest global structures resembling the mammalian visual system can emerge from simple learning rules.

Conclusions:

• Regular structures in artificial neural networks can evolve from learning rules driven by motion-based transformations.
• Motion-prediction error is a plausible mechanism for shaping visual system architecture.
• This work supports the idea that environmental dynamics play a critical role in the development of biological visual systems.