Multi-task and continual-learning capabilities of reservoir computing applied to speech recognition

The multitask- and continual-learning capabilities, which are both indispensable to living things coping with unpredictably changing environments, of single brainmorphic reservoir computing (RC) were investigated by applying it to spoken-word and speaker recognition tasks. By increasing the number of nodes and tuning the nonlinear recurrent dynamics at the edge of chaos, a single RC network achieved high scores for a simultaneous spoken-word and speaker recognition task, while a single LSTM, a representative recurrent neural network, could not balance the conflicting tasks and did not achieve the scores obtained in the case of processing each task separately. A new training scheme for RC is proposed to handle successive speaker recognition tasks. By storing time-averaged node activities of preceding tasks and loading the node activity history during additional task learning, both tasks were successfully identified without suffering the catastrophic forgetting observed in the case of not using a node activity history.