Learning to Stabilize Extreme Neural Machines with Metaplasticity

Training recurrent reservoir networks to solve complex tasks is a difficult problem due to extensive training times and the ambiguity of adjusting synaptic weights located deep within the architecture of the network. Recently, a novel model termed Extreme Neural Machine (ENM) has been proposed to perform one-shot learning in recurrent networks. A drawback of this approach, however, is that synaptic weights may become destabilized after training due to the random activation of recurrent units. In this paper, we propose a solution to this problem by incorporating metaplasticity in the learning rule of ENMs. With this novel approach, networks learned complex functions that can be recalled after a delay period. Using realworld data, networks were trained to produce sound waveforms obtained from spoken English words. Results show that metaplasticity improved task recall in recurrent neural networks.