Illustrative Discussion of MC-dropout Method in General Dataset: Uncertainty Estimation in Bitcoin.

Abstract

The past few years have witnessed the resurgence of uncertainty estimation in deep neural networks (DNNs). Providing uncertainty besides the predictions is desirable to provide a degree of belief about the predicted output in neural networks (NNs). Recent researches have introduced probabilistic approaches which are computationally less expensive than Bayesian neural networks (BNNs). Out of the existing approaches, we focus on probabilistic approach based Bayesian approximations known as Monte-Carlo dropout (MC-dropout). In this paper, we aim to provide an overview of the misconception arisen about MC-dropout, wherein criticism occurs. We fulfill our opinion using a 2-D synthetic dataset to derive insights about the empirical study. On the other hand, previous researches have often applied MC-dropout on classification tasks using image dataset. While, we provide an illustrative discussion of MC-dropout using a general dataset derived from Bitcoin blockchain known as Elliptic data. Using Elliptic data, we highlight the performance of uncertainty estimation using different sets of features. We further discuss the effect of MC-dropout regarding the uncertainty metrics when dealing with imbalanced data. The overall model have provided adequate results in terms of uncertainty measurements yielded by MC-dropout.