Continual learning of multiple cognitive functions with a brain-inspired temporal development mechanism

Area of Science:

• Artificial Intelligence
• Computational Neuroscience
• Machine Learning

Background:

• Current AI requires large networks for cognitive functions, unlike the brain's efficient, multi-functional learning.
• The brain utilizes temporal development mechanisms for knowledge transfer and redundancy prevention.
• Existing AI methods struggle with continual learning and energy efficiency.

Purpose of the Study:

• To propose a brain-inspired temporal development mechanism for continual learning in AI.
• To enable AI to learn multiple cognitive functions sequentially, from simple to complex tasks.
• To enhance AI's cognitive abilities while reducing network scale and energy consumption.

Main Methods:

• Developed a model simulating brain's cross-regional temporal development for AI.
• Implemented sequential evolution of long-range inter-module connections for knowledge transfer.
• Used feedback-guided local inhibition and pruning to eliminate task redundancies.

Main Results:

• Achieved continual learning of multiple cognitive functions without regularization, replay, or freezing.
• Demonstrated reduced network scale and energy consumption compared to direct learning.
• Attained superior accuracy on new tasks in cross-domain (PMI) and general datasets (CIFAR100, ImageNet).

Conclusions:

• Brain's developmental mechanisms offer a blueprint for biologically plausible, low-energy AI.
• The proposed method enables efficient continual learning and cognitive enhancement in AI systems.
• This approach provides a novel pathway for developing more sustainable and capable artificial intelligence.