Test Data Generation for Recurrent Neural Network Implementations
Katherine G. Skocelas, Byron DeVries · 2020
Despite their ubiquity, Artificial Neural Networks (ANNs) are still difficult to correctly implement and verify. While feedforward ANNs behave based on a clear cause and effect that maps an input to an output, Recurrent ANNs (RANNs) propagate internal (i.e., hidden layer) or external (i.e., output layer) values back to the inputs. Problematically, the benefits enabled by the recurrent links also introduce verification challenges. The cause and effect relationship of input to output units normally displayed in feedforward artificial neural networks may be altered due to internal values propagated forward in time via recurrent links. This verification difficulty is exacerbated in embedded systems (e.g., robotic or cyber-physical systems software) where in-house implementations of RANNs may be preferred, or even necessitated, due to hardware constraints. In this paper, we identify minimalistic test cases to verify the behavior of context units connected by internal recurrent links within RANNs.