Fig. 2: Protocol for training artificial neural networks with generative replay.

a On the first task, the main model [M] and a separate generator [G] are trained normally. When moving on to a new task, these two trained models first produce the samples to be replayed (see (b)). Those generated samples are then replayed along with the training data of the current task, and both models are trained further on this extended dataset. b To produce the samples to be replayed, inputs representative of the previous tasks are first sampled from the trained generator and then labelled based on the predictions made for them by the trained main model.