SWAP-GNG: A Scalable HW Architecture for Multi Growing Neural Gas Graph Learning for Continual Learning

In this paper, we propose a scalable hardware architecture for multiple Growing Neural Gas graph learning adapted for Continual Learning. By having a hardware architecture capable of capturing and learning continuously all the changes in the data distributions processed by a learning system, the design continuously adapts its behaviour to the data evolutions. The proposed architecture is described in details as well as the employed mechanisms for learning multiple graphs at the same time. The discussed work is validated through simulation in SystemC and described in VHDL for synthesis and performance evaluation purposes.