PhD thesis explores continual backpropagation to help neural networks keep learning

Richard Sutton shared the abstract of Shibhansh Dohare’s PhD thesis on plasticity, the ability of neural networks to learn new things. Dohare reports that some deep-learning systems lose plasticity as hidden units become dormant and similar to one another. He says continual backpropagation, developed with collaborators, reinitializes a small fraction of units at each step. In his studies, it maintained plasticity in many continual supervised-learning problems and helped reinforcement-learning algorithms recover from performance drops.