Model Compression for a Plasticity Neural Network in a Maze Exploration Scenario

Plastic neural networks provide a new type of biological heuristic method for “meta-learning” or “learning to learn”, opening a new door for research on lifelong learning and fast memory of artificial intelligence. They can combine typical neural networks with the famous Hebb’s rule in biological neurology, and uses Hebbian trace to express the connection strength between neurons. However, redundancy exists in previously designed plastic neural networks since not all neurons have strong connections. In this paper, we first time propose a model compression strategy for an RNN-based plastic neural network in a maze exploration scenario. With a reinforcement learning process, the network is trained and the hidden neurons have different variation trends. As convergent hidden neurons are conceptually able to memory some invariant feature of the maze, connections of nonconvergent hidden neurons are skillfully pruned. We successfully realize our approach and achieve plastic neural network compression in experiments. While ensuring the performance of the algorithm, the compression rate reaches 3.8 and the speedup rate is about 4.1 in theory and up to 16.