Logical and linear dependencies extraction from trained neural networks

The authors describe two methods of logical and linear dependencies extraction from trained neural networks. The main property of these methods is to interpret each unit as a logical-numerical or numerical operator. The logical-numerical (resp numerical) units perform the logical (resp numerical) task in the network. The first method relies on the architecture of the network, the second one favors theoretical modification of backpropagation (BP): two constraints are added to the standard BP to get logical-numerical operators and to help the network to learn discontinuous areas of functions. Simulation results on piecewise linear functions are reported