A probabilistic approach to object classification by neural trees

In this paper, a probabilistic approach is followed to improve the classification performances of neural trees. A neural tree whose nodes are generalized perceptrons without hidden layers and with activation function characterized by a sigmoidal behaviour is considered. The standard classification method may sometimes end up with wrong conclusions, e.g., the pattern is close to one of the decision hyperplanes. This situation occurs when the activation vector in one or more internal nodes (doubt nodes) of the NT is characterized by some values close to the highest value. To this end, multiple paths are followed by appropriately backtracking into the tree. A gain function is assigned to each node and a probabilistic technique to search for the path which maximizes this gain is proposed to improve the classification performances of the standard NT