Search

Proposes a new competitive learning algorithm that dynamically creates output neurons. The number of output neurons is increased as learning proceeds, whereas conventional competitive learning algorithms use all of the available output neurons during the entire learning phase. An acceptance test for the winning output neuron is performed using class thresholds if the number of created output neurons is less than the predefined maximum number. Accepted input vectors are used to adjust the reference vector of the winning output neuron. If an input vector is rejected, it is used as the initial reference vector of a new output neuron. The proposed method gets around the drawbacks of the conventional competitive learning algorithms by changing the class threshold values of output neurons dynamically. Experiments with remote sensing data and speech data indicate the superiority of the proposed algorithm in comparison to the conventional competitive learning methods. >

When two dimensional images are used as input to a neural network, the noise from the input device and small deformations in the end parts that occur in the processes of separating each pattern and size normalization lead to images shifted from the original learned image being input to the neural network which is a major cause of misrecognition. The inputs of the second order neural network (SONN) described consist of geometrically related nonlinear combinations of two pixels, and can be used for shift invariant pattern recognition. But the number of SONN input nodes increases in proportion to N/sup 2/, where N is the dimension of the input patterns, even if we only consider the shift invariance. Such large number of input nodes lead to slower learning aid recognition. In this paper, we propose a method for reducing the number of shift invariant SONN input nodes using combinations of input pattern pixels and the principle component analysis. Using the proposed method we are able to implement a shift invariant second order neural network with 2/5*N nodes. Due to the reduced number of input nodes, a 50% reduction in the learning and recognition time was obtained. >