GPOL: Gradient and Probabilistic approach for Object Localization to understand the working of CNNs

Convolutional neural networks have been a revolution in the field of Computer Vision and are being extensively used for the purpose of image classification, object detection, generation of captions etc. CNNs are mostly considered black boxes where the internal functioning is not known. The objective of this work is to provide an explanation of the functioning of the the predictions made by the CNN. We propose a new technique for comprehending the functioning of the middle layers of the neural network and the classifier operations. The proposed approach is capable of analyzing multifarious models which are trained for applications such as object detection and recognition. In this work, probabilistic approach and gradient based approach have been used for the purpose of object localization. Geometric mean of heatmaps of both the approaches has been done. In the former approach, the true object’s gradient’s are made to flow into the last convolutional layer for the purpose of determining the most significant points which would help to predict that particular object. In the probabilistic approach, CNN’s top down attention has been used which serves the purpose of generation of attention maps which are task specific. A probabilistic scheme (to select a significant neuron in the network) has been used during backpropagation of signals from top to down in the hierarchy of network. The proposed work has been executed on CLS-LOC dataset which is a part of Imagenet dataset. The proposed work is then compared with the previously developed techniques such as saliency maps, SmoothGrad, GradCam, Top Down Neural approach to exhibit the better accuracy of the proposed work.