Neural modularity helps organisms evolve to learn new skills without forgetting old skills

Area of Science:

• Artificial Intelligence
• Computational Neuroscience
• Machine Learning

Background:

• Artificial intelligence agents aim to learn diverse skills but face catastrophic forgetting, where new learning erases prior knowledge.
• This forgetting occurs as neural learning algorithms modify connections essential for previously acquired skills.

Purpose of the Study:

• To investigate if evolving modular neural networks can mitigate catastrophic forgetting.
• To explore how network organization, specifically modularity, impacts learning dynamics and skill retention.

Main Methods:

• Evolving neural networks with a cost for neural connections to induce modularity.
• Utilizing neuromodulation to selectively adjust learning rates based on environmental stimuli and rewards.
• Implementing separate modules for reinforcement learning and sensory processing.

Main Results:

• The connection cost technique successfully induced modularity in neural networks.
• Modular, sparsely connected networks demonstrated superior performance, learning new skills faster and retaining old skills better.
• A distinct reinforcement learning module contributed to improved learning efficiency.

Conclusions:

• Encouraging modularity in neural networks offers a promising solution to the problem of catastrophic forgetting.
• The inherent modularity observed in animal brains may serve to alleviate catastrophic forgetting, enhancing overall cognitive flexibility.