1: \begin{abstract}

2: Deep neural networks have become the default choice for many of the

3: machine learning tasks such as classification and regression. Dropout,

4: a method commonly used to improve the convergence of deep neural networks,

5: generates an ensemble of thinned networks with extensive weight sharing.

6: Recent studies that dropout can be viewed as an approximate variational

7: inference in Gaussian processes, and used as a practical tool to obtain

8: uncertainty estimates of the network. We propose a novel statistical

9: mechanics based framework to dropout and use this framework to propose

10: a new generic algorithm that focuses on estimates of the variance

11: of the loss as measured by the ensemble of thinned networks. Our approach

12: can be applied to a wide range of deep neural network architectures

13: and machine learning tasks. In classification, this algorithm allows

14: the generation of a don't-know answer to be generated, which can increase

15: the reliability of the classifier. Empirically we demonstrate state-of-the-art

16: AUC results on publicly available benchmarks.

17: \end{abstract}

18: