Back-propagation with diversive curiosity: An automatic conversion from search stagnation to exploration

This paper proposes a novel approach, namely, the Back-propagation with Diversive Curiosity (DCPROP) algorithm, for solving the “flat spot” problem and for escaping from local minima. Representing the diversive curiosity, an internal indicator is designed for BP algorithm, which detects the phenomenon of being trapped in local minima and the occurrence of premature convergence. Upon such detection, the neural network is activated again to explore optimal solution in search space and escape form local minima by means of stochastic disturbance. The proposed DCPROP algorithm is implemented and applied to two well-known face recognition problems, and the results are compared with Standard Back-propagation (SBP).