Stochastic Weight Perturbations Along the Hessian: A Plug-and-Play Method to Compute Uncertainty

Hariharan Ravishankar, Rohan Patil, Deepa Anand, Vanika Singhal, Utkarsh Agrawal, Rahul Venkataramani, Prasad Sudhakar

Abstract: An uncertainty score along with predictions of a deep learning model is necessary for acceptance and often mandatory to satisfy regulatory requirements. The predominant method to generating uncertainty scores is to utilize a Bayesian formulation of deep learning. In this paper, we present a plug-and-play method to obtain samples from an already optimized model. Specifically, we present a simple, albeit principled methodology, to generate a number of models by sampling along the eigen directions of the Hessian of the converged minimum. We demonstrate the utility of our methods on two challenging medical ultrasound imaging problems - cardiac view recognition and kidney segmentation.