A Memristor-Based Learning Engine for Synaptic Trace-Based Online Learning.

Area of Science:

• Neuromorphic Engineering
• Artificial Intelligence
• Materials Science

Background:

• Memristors are crucial for synaptic online learning in spiking neural networks (SNNs).
• Existing memristor approaches struggle with complex trace-based learning rules like STDP and BCPNN.
• There is a need for efficient hardware implementations of advanced SNN learning algorithms.

Purpose of the Study:

• To propose and realize a novel learning engine for trace-based online learning in SNNs.
• To enable the implementation of sophisticated learning rules, including STDP and BCPNN, using memristors.
• To demonstrate the energy efficiency and performance of the proposed learning engine.

Main Methods:

• Developed a hybrid learning engine combining memristor-based and analog computing blocks.
• Utilized memristors to emulate synaptic trace dynamics via their nonlinear physical properties.
• Integrated analog computing for essential operations like addition, multiplication, logarithm, and integration.
• Architected a reconfigurable engine to simulate both STDP and BCPNN learning rules.

Main Results:

• Achieved low energy consumption: 10.61 pJ/synaptic update for STDP and 51.49 pJ/synaptic update for BCPNN.
• Demonstrated significant energy reduction compared to ASIC counterparts (147.03× and 93.61× for 180 nm; 9.39× and 5.63× for 40 nm).
• Outperformed state-of-the-art neuromorphic platforms (Loihi, eBrainII) in energy efficiency by 11.31× and 13.13× for STDP and BCPNN, respectively.

Conclusions:

• The proposed learning engine effectively implements trace-based STDP and BCPNN learning rules in SNNs.
• This architecture offers substantial energy savings compared to existing ASIC and neuromorphic solutions.
• The memristor-based design paves the way for more efficient and powerful brain-inspired computing systems.