Scalable and Fully Configurable NoC-based Hardware Implemention of Growing Neural Gas for Continual Learning

Due to their accuracy and performances, neural networks (NN), like deep learning, are preferred in many application fields, but face challenges such as concept drift, requiring adaptable models for evolving data distributions. Concepts such as Continual (CL) or Incremental Learning (IL) are proposed as solutions to handle data changes over time and prevent catastrophic forgetting. Prototype-based methods have shown promise in addressing these challenges, offering robustness to adversarial attacks and interpretability. In this paper, we focus on the hardware implementation of one of the prototype-based methods, namely Growing Neural Gas, which is a dynamically evolving NN suited for CL. Its scalable and highly configurable architecture is presented in this work. The proposed architecture has been validated through simulations in SystemC framework demonstrating the proof of concept for CL. Additionally, the results of hardware implementation targeting FPGA are presented, illustrating the resource utilization of its major components.