Classifier-memory-based parameter adaptation

Abstract

There is a phenomenon named catastrophic forgetting, where a neural network forgets information learned from past tasks when being trained for new tasks. Memory-based parameter adaptation (MbPA) has been suggested to address this issue recently by augmenting a neural network with non-parametric, episodic memory. In this framework, trained instances are saved to an experience memory and parameters of the neural network are adapted to some examples stored in memory. The paper has shown MbPA performing a Maximum A Posteriori estimation and attention mechanism is a certain kind of MbPA. In this thesis, we try to improve robustness and complexity of MbPA. Here, we focus on an idea of training a classifier which classifies each problem's input domain. Then we can filter the candidate of context for parameter adaptation to the train set of classified problem. The suggested algorithm, classifier-memory-based parameter adaption, has shown experimental results with improved complexity and comparable (or even better in some setting) performance with MbPA.