Accretionary Learning With Deep Neural Networks With Applications

One of the fundamental limitations of Deep Neural Networks (DNN) is their inability to acquire and accumulate new cognitive capabilities in an incremental or progressive manner. When data appear from object classes not among the learned ones, a conventional DNN would not be able to recognize them due to the fundamental formulation that it assumes. A typical solution is to re-design and re-learn a new network, most likely an expanded one, for the expanded set of object classes. This process is quite different from that of a human learner. In this paper, we propose a new learning method named Accretionary Learning (AL) to emulate human learning, in that the set of object classes to be recognized need not be fixed, meaning it can grow as the situation arises without requiring an entire redesign of the system. The proposed learning structure is modularized, and can dynamically expand to learn and register new knowledge, as the set of objects grows in size. AL does not forget previous knowledge when learning new data classes. We show that the structure and its learning methodology lead to a system that can grow to cope with increased cognitive complexity while providing stable and superior overall performance.