Incremental learning using hidden layer activations-Tests on fish identification data

The multilayer perceptron model for pattern recognition problems does not adapt easily when the problem environment is modified slightly by adding a new output category. The standard solution is to discard the old network and train a new one. The use of hidden layer activations to retrain the old network incrementally is explored. The results of an empirical evaluation of the method on a fish identification problem are quite encouraging. The performance of the incrementally trained networks compares very well with that of a new network, while the training effort is smaller by a factor of five