Unsupervised Neuron Selection for Mitigating Catastrophic Forgetting in Neural Networks

Catastrophic forgetting is a well studied problem in artificial neural networks in which past representations are rapidly lost as new representations are constructed. We hypothesize that such forgetting occurs due to overlap in the hidden layers, as well as the global nature in which neurons encode information. We introduce a novel technique to mitigate forgetting which effectively minimizes activation overlapping by using online clustering to effectively select neurons in the feedforward and back-propagation phases. We demonstrate the memory retention properties of the proposed scheme using the MNIST digit recognition data set.