SynapNet: A Complementary Learning System Inspired Algorithm With Real-Time Application in Multimodal Perception.

Area of Science:

• Artificial Intelligence
• Neuroscience
• Machine Learning

Background:

• Catastrophic forgetting is a major challenge in continual learning (CL), where neural networks lose previously acquired knowledge when learning new tasks.
• The mammalian brain mitigates this by consolidating memories across the hippocampus and neocortex.
• Existing CL methods often struggle to balance plasticity and stability, leading to performance degradation.

Purpose of the Study:

• To propose a novel CL framework inspired by mammalian memory consolidation to overcome catastrophic forgetting.
• To enhance neural network performance in incremental learning scenarios.
• To demonstrate the practical applicability of the framework in real-world robotic tasks.

Main Methods:

• A dual-model approach combining a plastic (hippocampal-like) and a stable (neocortical-like) component.
• Integration of a variational autoencoder (VAE) for pseudo-rehearsal.
• Application of lateral inhibition masks and a sleep phase for gradient regularization and representation reorganization.

Main Results:

• Empirical evaluations on class-incremental and domain-incremental datasets show significant performance improvements compared to standard benchmarks.
• The framework successfully demonstrated real-time incremental object classification in a physical environment using a soft pneumatic gripper.
• Significant backward knowledge transfer (KT) was observed, indicating effective knowledge retention and utilization.

Conclusions:

• The proposed brain-inspired CL framework effectively mitigates catastrophic forgetting.
• The combination of dual models, VAE pseudo-memory, lateral inhibition, and sleep phases enhances learning stability and performance.
• The framework shows promise for practical applications in robotics and real-time incremental learning systems.