Sparsity through evolutionary pruning prevents neuronal networks from overfitting.

Area of Science:

• Neuroscience
• Artificial Intelligence
• Computational Biology

Background:

• Modern machine learning (ML) leverages increasing computational power, leading to models with vast numbers of parameters.
• Despite advancements, ML models lack the general intelligence of the human brain, failing to solve diverse tasks with a single architecture.
• Biological neural networks possess inherent hierarchical structures, unlike ML models trained from scratch.

Purpose of the Study:

• To investigate the structural basis of biological neural networks using a bottom-up approach.
• To explore how evolutionary optimization and network structure impact generalization performance in artificial neural networks.
• To determine if sparsity is a key factor in achieving better generalization in machine learning.

Main Methods:

• Evolutionary training of small neural networks on a maze task.
• Implementing dynamic decision-making with delayed rewards within the task.
• Comparing the generalization performance of sparsely connected networks against fully connected networks.

Main Results:

• Evolutionary optimization of neural networks for a maze task was achieved.
• Random severance of connections during training led to improved generalization performance.
• Sparsely connected networks outperformed fully connected networks in generalization.

Conclusions:

• Sparsity is a crucial property for effective neural network generalization.
• The findings suggest that incorporating sparsity principles could enhance modern machine learning approaches.
• This study provides insights into the structural underpinnings of biological neural intelligence.