Quantification and Analysis of Layer-wise and Pixel-wise Information Discarding
This provides a new perspective for explaining DNNs, focusing on information discarding in a pixel-wise manner, which is incremental compared to existing information bottleneck theory.
The paper tackles the problem of explaining how input information is discarded during forward propagation in deep neural networks by defining entropy-based metrics for pixel-wise and layer-wise analysis, showing connections to DNN performance in experiments.
This paper presents a method to explain how the information of each input variable is gradually discarded during the forward propagation in a deep neural network (DNN), which provides new perspectives to explain DNNs. We define two types of entropy-based metrics, i.e. (1) the discarding of pixel-wise information used in the forward propagation, and (2) the uncertainty of the input reconstruction, to measure input information contained by a specific layer from two perspectives. Unlike previous attribution metrics, the proposed metrics ensure the fairness of comparisons between different layers of different DNNs. We can use these metrics to analyze the efficiency of information processing in DNNs, which exhibits strong connections to the performance of DNNs. We analyze information discarding in a pixel-wise manner, which is different from the information bottleneck theory measuring feature information w.r.t. the sample distribution. Experiments have shown the effectiveness of our metrics in analyzing classic DNNs and explaining existing deep-learning techniques.