Reinforcement learning techniques for network-based transfer learning

Abstract

Embodiments provide reinforcement learning (RL) techniques for transferring machine learning obtained in one domain to another. A neural network (NN) of a source domain may be trained using training data associated with that domain. In some cases, the training data is inaccessible to a target domain for which a similar NN is desired. A number of target NNs may be initialized with a portion of the parameters transferred from the source NN. A computing agent may utilize RL techniques to train the target NNs using training data of the target domain. The target NN with the most beneficial subset of transferred parameters may be selected. By using the RL techniques to identify the most advantageous combination of transferred parameters, a similarly accurate NN may be provided in the target domain at a fraction of the otherwise needed time and without accessing the original training data.