Abstract :The trade-off between interpretability and accuracy in machine learning is a long-standing challenge in Machine learning. The rise of deep neural networks and their ubiquitous usage has brought this problem to the forefront once again. The power of neural networks comes mainly from its multi-layer feature representation through the use of "hidden nodes". By their very nature, hidden nodes are not interpretable, and can at best be conjectured to represent some latent factor that aids in the final task, e.g. a hidden node may detect eyes in an image, and later layers can use this information for predicting the presence of a cat. Hence an entire class of approaches that try to "understand" a learned network based on various "visualizations" have arisen. This "understanding" can be used by the programmer for debugging the learned network or conveyed to the end-user as a reason for a particular prediction. In this talk, we will discuss some standard visualization approaches used for understanding the predictions of deep networks on images, speech, and audio tasks.
Speaker:Saurabh Desai Video
Talk Title: Ablation-CAM:Visual Explanations for Deep Convolutional Network via Gradient-free Localization
Bio: Saurabh is currently a graduate student at Oregon State University under Dr.Stefan Lee in the Computer Science department. Previously, he used to work as a Project Associate at RBCDSAI, IIT Madras under supervision of Dr. Balaraman Ravindran and Dr.Harish Guruprasad. His interests lie in Explainable AI, Computer Vision, Reinforcement learning and Embodied AI.
Abstract :
In response to recent criticism of gradient-based visualization techniques, we propose a new methodology to generate visual explanations for deep Convolutional Neural Networks (CNN) - based models. Our approach - Ablation-based Class Activation Mapping (Ablation CAM) uses ablation analysis to determine the importance (weights) of individual feature map units w.r.t. class. Further, this is used to produce a coarse localization map highlighting the important regions in the image for predicting the concept. Our objective and subjective evaluations show that this gradient-free approach works better than state-of-the-art Grad-CAM technique. Moreover, further experiments are carried out to show that Ablation-CAM is class discriminative as well as can be used to evaluate trust in a model.