Fig. 4
From: CLeaR: An adaptive continual learning framework for regression tasks
The CLeaR instance contains an autoencoder and a fully-connected neural network in the block Models. The MSE(X, \(\hat {X}\)) is compared to Threshold_a to detect the change of P(X), and the MSE(Y, \(\hat {Y}\)) is compared to Threshold_p to detect the change of P(Y|X). The novelty buffer in Buffers_a is for updating the autoencoder and the novelty buffer in Buffers_p is for updating the predictor. Models are evaluated with the data in the familiarity buffer