Extended tower algorithm for multi category classification
T.N. Nagabhushana, S. K. Padma, Maithri Bairy · 2004
This paper presents an extension to the constructive learning tower algorithm for multicategory classification. The tower algorithm proposed by Stephen Gallant (Stephen I Gallant, (1990)) is one of the constructive learning algorithms for building a neural network during the training phase. Constructive learning algorithms are shown to be more robust since they build optimal neural network architectures. Gallant's tower algorithm basically performs two category classifications. Extensions to the two category tower algorithm for multicategory classification is straightward. It can be accomplished by adding a layer of M neurons each time a new layer is added to the tower. The constructive tower algorithm develops an architecture which could form the basis for incremental learning where new data is learnt without forgetting the prior knowledge. This feature makes it suitable for on-line applications. Different variations of the tower algorithm for multicategory classification are presented and observations discussed.