PLASTICITY-AWARE SPIKING TRANSFORMERS: A NEUROCOMPUTATIONAL FRAMEWORK FOR LOW-POWER CONTINUAL LEARNING IN DYNAMIC ENVIRONMENTS

Real-time intelligent systems operating in dynamic environments require the capacity for continual learning, low power consumption, and biologically inspired adaptability. This paper introduces Plasticity-Aware Spiking Transformers (PAST)—a neurocomputational architecture that fuses spiking neural networks (SNNs) with temporal attention transformers, enriched by plasticity-driven learning and adaptive memory modules. Designed for deployment on neuromorphic hardware, PAST models mimic the human brain's energy-efficient mechanisms to perform continual learning with minimal catastrophic forgetting. Experimental results show that PAST achieves significant improvements in accuracy, energy consumption, and adaptation compared to conventional deep learning models across dynamic benchmarks like Split-CIFAR, DVS Gesture, and Permuted MNIST.