Genesis: A Neuromorphic Chip with Lifelong Learning on-Device

Deep neural networks (DNNs) have been proven to excel in various domains, from image classification to natural language processing. Despite their remarkable performance, these models face a significant limitation--lack of lifelong learning capabilities, impeding their capacity to assimilate new knowledge while retaining previously acquired information. Addressing this challenge proves even more daunting when adapting these capabilities to resource-constrained edge platforms with stringent size, weight, area, and power (SWaP) constraints. This project is dedicated to crafting a resilient lifelong learning agent tailored for operation in such constrained environments. Two state-of-the-art hardware-friendly lifelong learning models, TACOS and MetaplasticNet, were developed, drawing inspiration from neuro-inspired mechanisms like Metaplasticity and synaptic consolidation. In tandem with algorithmic innovations, we introduced two cutting-edge CMOS/Memristor hybrid accelerators--Genesis-v1 and Genesis-v2. These accelerators not only execute lifelong learning tasks in real-time at approximately 20-30 frames per second but also operate with exceptional efficiency, consuming a minimal power of 20mW. This integrated approach pushes the boundaries of lifelong learning in DNNs and provides a robust solution for on-device learning in resource-constrained environments.