Evolving a Dynamic Predictive Coding Mechanism for Novelty Detection

Novelty detection is a machine learning technique which identifies new or unknown information in data sets. We present our current work on the construction of a new novelty detector based on a dynamical version of predictive coding. We compare three evolutionary algorithms, a simple genetic algorithm, NEAT and FS-NEAT, for the task of optimising the structure of an illustrative dynamic predictive coding neural network to improve its performance over stimuli from a number of artificially generated visual environments. We find that NEAT performs more reliably than the other two algorithms in this task and evolves the network with the highest fitness. However, both NEAT and FS-NEAT fail to evolve a network with a significantly higher fitness than the best network evolved by the simple genetic algorithm. The best network evolved demonstrates a more consistent performance over a broader range of inputs than the original network. We also examine the robustness of this network to noise and find that it handles low levels reasonably well, but is outperformed by the illustrative network when the level of noise is increased.