Uncertainty-based modulation for lifelong learning.

Area of Science:

• Artificial Intelligence
• Computational Neuroscience
• Machine Learning

Background:

• Continuous, lifelong learning is crucial for intelligent agents in dynamic environments.
• Existing algorithms often suffer from catastrophic forgetting.

Purpose of the Study:

• To present a novel algorithm inspired by human brain neuromodulatory mechanisms for continuous lifelong learning.
• To integrate and expand upon Adaptive Resonance Theory (ART) proposals.

Main Methods:

• Developed a new algorithm incorporating "neuromodulatory" mechanisms for continuous, self-supervised, and one-shot learning.
• Evaluated algorithm components in benchmark experiments and integrated into a simulated drone agent for closed-loop learning.
• Utilized high-dimensional inputs from state-of-the-art detection and feature extraction algorithms.

Main Results:

• Demonstrated stable learning without catastrophic forgetting in benchmark experiments.
• Achieved high classification accuracy (>94%) for continuous learning of new tasks and under changed conditions in a virtual environment.
• Showcased the algorithm's flexibility and adaptability in a simulated embodied agent.

Conclusions:

• The proposed algorithm enables robust lifelong learning in intelligent agents without catastrophic forgetting.
• Closed-loop development, where environment and agent behavior guide learning, is critical for effective continuous learning.
• The algorithm is a flexible component for existing AI systems, with future work focusing on knowledge-seeking and broader neuromodulation.