Dual memory model for experience-once task-incremental lifelong learning.

Area of Science:

• Computational Neuroscience
• Artificial Intelligence
• Machine Learning

Background:

• Experience replay (ER) is a key neuroscience-inspired technique for lifelong learning.
• Existing ER methods use simplified memory modules, limiting their effectiveness.
• Neuroscience suggests fine-grained memory systems, like the PFC-HC dual-store circuit, offer greater potential.

Purpose of the Study:

• To computationally model the memorization, consolidation, and rehearsal processes of the PFC-HC dual-store memory circuit.
• To introduce the Dual-Memory (Dual-MEM) model for more effective lifelong learning.
• To address limitations in current experience replay mechanisms.

Main Methods:

• Developed the Dual-Memory (Dual-MEM) model, incorporating short-term memory for current tasks and consolidated long-term memory for future rehearsal.
• Proposed two optimization learning policies: Direct Retrieval Learning and Mixup Retrieval Learning, simulating different memory retrieval strategies.
• Evaluated Dual-MEM on eight benchmark datasets under an experience-once setting.

Main Results:

• Dual-MEM demonstrated compelling performance across multiple benchmarks.
• The model achieved high learning and memory utilization efficiencies.
• Effectiveness was shown even in the challenging experience-once learning scenario.

Conclusions:

• The Dual-MEM model successfully emulates the PFC-HC dual-store memory system for enhanced lifelong learning.
• Dual-MEM offers a more sophisticated approach to memory replay compared to existing methods.
• The model provides an efficient and effective solution for experience-once learning challenges.