Neural Criticality Metric for Object Detection Deep Neural Networks

The complexity of state-of-the-art Deep Neural Network (DNN) architectures exacerbates the search for safety relevant metrics and methods that could be used for functional safety assessments. In this article, we investigate Neurons' Criticality (the ability to affect the decision process) for several object detection DNN architectures. As a first step, we introduce the Neural Criticality metric for object detection DNNs and set a theoretical background. Subsequently, by conducting experiments, we verify that removing one neuron from the computational graph of a DNN can have a significant (positive, as well as negative) influence on the prediction's precision (object classification and localization). Finally, we build statistics for each neuron from pre-trained networks on the COCO object detection validation dataset and examine the network stability for the most critical neurons in order to prove our metric's validity