HWA: Hyperparameters Weight Averaging in Bayesian Neural Networks

Belhal Karimi, Ping Li

Keywords: Bayesian, neural, network, averaging, hyperparameters, generalization

TL;DR: We propose an averaging procedure, namely HWA, for training Bayesian Neural Networks, leading to an improvement of empirical test accuracy.

Abstract:

Bayesian neural networks attempt to combine the strong predictive performance of neural networks with formal quantification of uncertainty of the predicted output in the Bayesian framework.
In deterministic deep neural network, the confidence of the model and the predictions at inference time are left alone.
Applying randomness and Bayes Rule to the weights of a deep neural network is a step towards achieving this goal.
Current state of the art optimization methods for training Bayesian Neural Networks are relatively slow and inefficient, compared to their deterministic counterparts.
In this paper, we propose HWA (Hyperparameters Weight Averaging) algorithm that exploits an averaging procedure in order to optimize faster and achieve better accuracy.
We develop our main algorithm using the simple averaging heuristic and demonstrate its effectiveness on the space of the hyperparameters of the networks random weights.
Numerical applications are presented to confirm the empirical benefits of our method.