COSC490 Project: Pseudorehearsal
Abstract
In artificial neural networks, catastrophic forgetting is where learning a new item has a catastrophic detrimental effect on previous learning. Pseudorehearsal is a rehearsal technique that allows synaptic plasticity while minimising catastrophic forgetting, without needing to store previously learned items; this report reviews it in simple MLPs and modern deep networks, and investigates it empirically as layers are added.