Automatic Test Generation Using Neural Networks

We describe a new automatic test pattern generation (ATPG) methodology that has the potential to exploit fine-grain parallel computing and relaxation techniques. Our approach is radically different from the conventional methods used to generate tests for circuits from their gate level descriptions. A digital circuit is represented as a bidirectional network of neurons. The circuit function is coded in the firing thresholds of neurons and the weights of interconnection links. This neural network is suitably reconfigured for solving the ATPG problem. A fault is injected into the neural network and an energy function is constructed with global minima at test vectors. Global minima are determined by a probabilistic relaxation technique augmented by a directed search. Preliminary results on combinational circuits confirm the feasibility of this technique.