WeightUncertainty
PyTorch implemenation of Weight Uncertainty in Neural Networks.
This repository provides an implementation of the Bayes-by-Backprop
framework as described in the Weight Uncertainty in Neural Networks
paper. The code provides a simple PyTorch interface and has been
designed to be extensible, allowing for implementation of additional
functionality such as alternate priors or variational posteriors. In
addition, modules that employ the Local Reparameterisation Trick have
been implemented to provide the opporunity for faster, stable training.
- [x] Python 3.6+
- [x] MIT License
Overview:
The Weight Uncertainty in Neural Networks
(WUINN) paper provides a framework known as Bayes-by-Backprop which
allows for learning a probability distribution on the weights of a
Neural Network. The network weights are regularised by minimising the
ELBO cost given a prior / posterior distribution, this helps avoid the
common pitfalls of conventional Neural Networks where overfitting is a
major concern and predictions are skewed by an inability to correctly
assess uncertainty in the training data.
The authors of WUINN utilise a Scale Mixture prior and a Gaussian posterior. This combination is computationally intractable and, as such, a variational approximation to the posterior must be found - this is completed through Monte Carlo sampling. Due to the sampling method employed, the parameters of the posterior distribution are still capable of being trained through standard gradient descent. Section 3.2 in the paper provides further information on this.
The minibatching / KL-reweighting scheme mentioned in the paper has been implemented to ensure that the first few minibatches are heavily influenced by the complexity cost and later batches are predominantly influenced by the data. The weights used are simply a function of the current batch index and the total number of batches.
This work has been extended through the use of a Gaussian distribution for the prior and posterior - this allows the local reparameterisation trick (LRT) to be used. As the LRT samples activations rather than the weights, the KL Divergence can not be calculated through the use of MC sampling; instead the closed form must be used. Fortunately this is possible for the Gaussian prior / posterior combination.
Example:
An example of a Bayes-by-Backprop network has been implemented in
mnist_bnn.py - this example can be run with:
python3 mnist_bnn.py
References:
@misc{blundell2015weight,
title={Weight Uncertainty in Neural Networks},
author={Charles Blundell and Julien Cornebise and Koray Kavukcuoglu and Daan Wierstra},
year={2015},
eprint={1505.05424},
archivePrefix={arXiv},
primaryClass={stat.ML}
}