Unsupervised adaptation to improve fault tolerance of neural network classifiers
A. Nugent, Garrett T. Kenyon, Reid B. Porter · 2004
We investigate how to exploit the dynamics of unsupervised online learning rules for fault tolerance in neural network classifiers. We first design an adaptation mechanism that keeps neural network weights at a useful fixed point for classification problems. We then demonstrate the robustness of the system when the network inputs are subjected to faults.