Haozhe Shan

A two-system RNN achieves continual learning by using a 'what' system to infer compositional structures in tasks and a 'how' system to compose low-rank components, demonstrating transfer and compositional generalization.