A modified s neuron and its application to scale-invariant classification

摘要

S neuron, a scale-invariant neuron, is successfully adopted in Neocognitron to function as a local feature extractor. But, in operation its weights may be updated unboundedly. In addition, its scale-invariant property will not be held when the learning patterns are noisy. In this paper, a Modified S Neuron (MSN) and its novel learning rule are introduced to overcome the drawbacks of S neuron. It is shown that after an MSN learns a specific pattern set the mean of its excitatory weight converges to the scaled mean of the learned pattern set and its inhibitory weight is equal to h norm of the excitatory weight. Moreover, by applying MSNs a self-organizing scale-invariant classifier which is a two-layer structure with fully-connected weights is constructed. Two learning algorithms are given for training the classifier. It is demonstrated that the classification rate can be improved, especially when the selectivity of MSN becomes adaptive. Finally, a simple experimental simulation is given for verification.