Autonomous evolution of topographic regularities in artificial neural networks

Area of Science:

• Artificial Intelligence
• Computational Neuroscience
• Evolutionary Computation

Background:

• Neuroevolution (NE) has evolved artificial neural networks (ANNs) for over 25 years.
• Current NE algorithms produce ANNs lacking biological brain characteristics, limiting mainstream adoption.
• A gap exists in understanding how to imbue evolved ANNs with biological brain-like features.

Purpose of the Study:

• To investigate if introducing geometric properties to evolved ANNs can yield brain-like characteristics.
• To explore the impact of spatial coordinates on ANNs evolved via neuroevolution.
• To determine if evolved ANNs can develop topographic maps and improve generalization.

Main Methods:

• Utilized the hypercube-based neuroevolution of augmenting topologies (NEAT) algorithm.
• Introduced spatial coordinates (locations) to neurons within the evolved ANNs.
• Conducted experiments evolving ANNs for the game of checkers.

Main Results:

• Evolved ANNs spontaneously developed topographic maps with symmetries and regularities.
• The ability to evolve topographic maps provided a significant advantage in generalization.
• Analysis revealed a correlation between player generality and smoother, more contiguous connectivity patterns.

Conclusions:

• Introducing geometry to NE algorithms allows ANNs to acquire brain-like characteristics.
• Evolved topographic maps are crucial for improved generalization in ANNs.
• The study suggests NE algorithms can evolve increasingly relevant ANNs for neural computation research.