A hierarchical incremental learning approach to task decomposition

Abstract

In this paper, we propose a new task decomposition approach- hierarchical incremental class learning (HICL). In this approach, a K -class problem is divided into K sub-problems. The sub-problems are learnt sequentially in a hierarchical structure with K sub-networks. Each sub-network takes the output from the sub-network immediately below it as well as the original input as its input. The output from each sub-network contains one more class than the sub-network immediately below it, and this output is fed into the sub-network above it. It not only reduces harmful interference among hidden layers, but also facilitates information transfer between classes during training. The later sub-networks can obtain learnt information from the earlier sub-networks. We also proposed two ordering algorithms - Minimal-Side-Effect-First ordering method based on Class Decomposition Error (MSEF-CDE) and Minimal Side-Effect Ordering based on Fisher's Linear Discriminant (MSEF-FLD) to determine the hierarchical relationship between the sub-networks. The proposed HICL approach shows smaller regression error and classification error than classical decomposition approaches.