Oscillatory training: cooperatively training multiple tasks in physical systems without batching

We focus on training numerous independent tasks simultaneously in the same physical system. Typically, when training multiple tasks, each task is either learned sequentially, or its corresponding learning update is stored in memory before an external supervisor provides a batched learning. Here, we introduce a methodology for simultaneous training of multiple tasks without explicit, memory-based batching. We input all data simultaneously through oscillatory external driving at incommensurate frequencies. The response of the network to each individual drive signal may then be separately extracted through frequency-resolved readouts. By treating each oscillatory mode as an independent communication channel, we accomplish massively parallel, in situ training. This cooperative learning protocol, where the same physical system is simultaneously trained to perform many tasks, enables temporal speed-ups and precise external control. We show that in some cases, because the tasks are learned cooperatively, solutions can be found that were not accessible by sequential training.