Incremental learning in terms of output attributes

Abstract

This paper deals with the situation where outputs are introduced to a neural network incrementally. Conventionally, when new outputs are introduced to a neural network, the old network would be discarded and a new network would be retrained to integrate the old knowledge with the new knowledge. However, this method is likely to waste much computation time due to the loss of learnt knowledge in the existing network. As such, how to integrate both old and new knowledge to form a single structure solution is our primary interest. In this paper, we present three Incremental Output Learning (IOL) algorithms for incremental output learning. When a new output is introduced to the original problem, a new sub-network is trained under IOL to acquire the new knowledge and the outputs from the new sub-network are integrated with the outputs of the existing network. The experimental results from several benchmarking datasets show that our methods are more effective and efficient than retraining.