Long Timescale Credit Assignment in NeuralNetworks with External Memory

Neural networks with external memories (NNEM),such as the Neural Turing Machine [1] and Memory Networks [2], have often been compared to the human hippocampus in their ability to maintain episodic information over long timescales [3]. But so far, these networks have only been trained on tasks requiring memory storage comparable to a few minutes, whereas the hippocampus stores information on the order of years. It is well known that this sort of long-term episodic storage is vital for overcoming the problem of catastrophic interference [4], which is central to the challenge of continual learning.