Adaptive Memory-Enhanced Time Delay Reservoir and its Memristive Implementation

The Time Delay Reservoir (TDR) is a hardware-friendly machine learning approach that reduces connection overhead in neural networks and can be implemented in various systems. However, it struggles with long-term dependency tasks. We introduce a higher-order delay unit to enhance reservoir memory, using Particle Swarm Optimization for adaptivity. Our experiments show superior performance over seven existing methods for both short- and long-term memory datasets. We also propose a memristive implementation of TDR, using dynamic memristors and memristor-based delay elements, which is feasible, effective, and more efficient in terms of circuit area and power consumption compared to traditional hardware reservoirs.