Knowledge-based cascade-correlation: using knowledge to speed learning

Research with neural networks typically ignores the role of knowledge in le arning by initializing the network with random connection weights. We exami ne a new extension of a well-known generative algorithm, cascade-correlatio n. Ordinary cascade-correlation constructs its own network topology by recr uiting new hidden units as needed to reduce network error. The extended alg orithm, knowledge-based cascade-correlation (KBCC), recruits previously lea rned sub-networks as well as single hidden units. This paper describes KBCC and assesses its performance on a series of small, but clear problems invo lving discrimination between two classes. The target class is distributed a s a simple geometric figure. Relevant source knowledge consists of various linear transformations of the target distribution. KBCC is observed to find , adapt and use its relevant knowledge to speed learning significantly.