Overcoming Catastrophic Interference by Conceptors
This addresses catastrophic forgetting in deep feedforward networks for continual learning, but it is incremental as it extends existing conceptor methods from reservoir computing to deep networks.
The paper tackles catastrophic interference in continual learning by proposing a variant of back-propagation called conceptor-aided back-prop (CAB), which uses conceptors to shield gradients and prevent degradation of previously learned tasks, and reports that CAB outperforms two other methods on the disjoint MNIST task.
Catastrophic interference has been a major roadblock in the research of continual learning. Here we propose a variant of the back-propagation algorithm, "conceptor-aided back-prop" (CAB), in which gradients are shielded by conceptors against degradation of previously learned tasks. Conceptors have their origin in reservoir computing, where they have been previously shown to overcome catastrophic forgetting. CAB extends these results to deep feedforward networks. On the disjoint MNIST task CAB outperforms two other methods for coping with catastrophic interference that have recently been proposed in the deep learning field.