Penalizing Small Errors Using an Adaptive Logarithmic Loss

Loss functions are error metrics that quantify the difference between a prediction and its corresponding ground truth. Fundamentally, they define functional landscape for traversal by gradient descent. Although numerous loss have been proposed to date in order handle various machine learning problems, little attention has given enhancing these better traverse landscape. In this paper, we simultaneously significantly mitigate two prominent problems medical image segmentation namely: i) class imbalance foreground background pixels ii) poor function convergence. To end, propose an Adaptive Logarithmic (ALL) function. We compare with existing state-of-the-art on ISIC 2018 dataset, nuclei dataset as well DRIVE retinal vessel dataset. measure performance of our methodology benchmark demonstrate performance. More generally, show system can be used framework training deep neural networks.