Alleviating catastrophic forgetting using context-dependent gating and synaptic stabilization.

Area of Science:

• Artificial Intelligence
• Neuroscience
• Machine Learning

Background:

• Artificial neural networks (ANNs) suffer from catastrophic forgetting, where learning new tasks erases previously acquired knowledge.
• This forgetting occurs because training on new tasks alters crucial connection weights essential for prior tasks.
• Existing methods focus on stabilizing important connection weights to mitigate this issue.

Purpose of the Study:

• To introduce a novel method to prevent catastrophic forgetting in ANNs.
• To enhance the continual learning capabilities of ANNs by drawing inspiration from biological learning mechanisms.
• To develop a computationally efficient and easily implementable solution for lifelong learning in artificial intelligence.

Main Methods:

• Proposed a context-dependent gating signal to activate sparse, non-overlapping patterns of units for each task.
• Integrated this gating mechanism with existing weight stabilization techniques.
• Tested the method on both feedforward and recurrent network architectures using supervised and reinforcement learning.

Main Results:

• The proposed gating method significantly reduced catastrophic forgetting in ANNs.
• ANNs maintained high performance across a large number of sequentially presented tasks.
• The method proved effective for various network architectures and learning paradigms.
• Combining gating with weight stabilization yielded superior results.

Conclusions:

• A context-dependent gating signal is an effective complementary strategy to combat catastrophic forgetting in ANNs.
• This approach, inspired by in vivo mechanisms, enhances continual learning and lifelong performance.
• Employing multiple, complementary methods, similar to brain function, is a powerful strategy for robust continual learning.