Non-Linear behavior of CNN model interpretation using Saliency Map
Sudhanshu Saurabh, P. K. Gupta · 2022
The complex behavior of the network limits human perception of the network’s success, which is among the major challenges of the convolutional neural network. The saliency map is an useful approach for analyzing and visualizing the non -linear behavior of a feed forward network. Since the DNN model is nonlinear, the interpretation is not straightforward. We intro-duce an interpretation method that observes the saliency map and visual input, which is used to calculate the component-wise product to remove the noise. Using gradient-based visualization, multimodal convolutional networks perform for visual inputs. The component-wise product is used to visualize the input by analyzing the learned weights of the layers. The saliency map is generated by the guided back-propagation of the ReLU activation function. In this paper, we aim at a popular branch of description techniques, often referred to as saliency methods. Our main objective is to find input features such as image pixels.