Real-time activation pattern monitoring and uncertainty characterisation in image classification

Deep neural networks (DNNs) have become very popular recently and have proven their potential especially for image classification. However, their performance depends significantly on the network structure and data quality. This paper investigates the performance of DNNs and especially of faster region based convolutional neural networks (R-CNNs), called faster R-CNN when the network testing data differ significantly from the training data. This paper proposes a framework for monitoring the neuron patterns within a faster RCNN by representing distributions of neuron activation patterns and by calculating corresponding distances between them, with the Kullback-Leibler divergence. The patterns of the activation states of ‘neurons’ within the network can therefore be observed if the faster R-CNN is ‘outside the comfort zone’, mostly when it works with noisy data and data that are significantly different from those used in the training stage. The validation is performed on publicly available datasets: MNIST [1] and PASCAL [2] and demonstrates that the proposed framework