SENSITIVITIES TO INPUTSMultilayer feedforward networks are often used for modeling complex relationships between the data sets. Deleting unimportant data components in the training sets could lead to smaller networks and reduced-size data vectors. This can be achieved by analyzing the total disturbance of network outputs due to perturbed inputs. The search for redundant data components is performed for networks with continuous outputs and is based on the concept in sensitivity of linearized neural networks. The formalized criteria and algorithm for pruning data vectors are formulated and illustrated with examples.

No abstract

No abstract

No abstract

Several neural network architectures for computing parity problems are described. Feedforward networks with one hidden layer require N neurons in the hidden layer. If fully connected feedforward networks are considered, the number of neurons in the hidden layer is reduced to N/2. In the case of fully connected networks with neurons connected in cascade, the minimum number of hidden neurons is between logl(N+I)-l and logz(N+l). This paper also describes hybrid neuron architectures with linear and threshold-like activation functions. These hybrid architectures require the least number of weights. The described architectures are suitable for hardware implementation since the majority of weights equal +I and weight multiplication is not required. The simplest network stmctures are pipeline architectures where all neurons and their weights are identical. All presented architectures and equations were verified with MATLAB code for parity-N problems as large as N=100.

No abstract

This paper introduces a new approach to inverse control. Unlike using commonly known method ofplant inverse dynamics learning, the control sequence is calculated using inverse mapping approach. Both methods are compared for nonlinear plant examples and selected desired waveforms.

No abstract