Implicit Weight Uncertainty in Neural Networks

This repository contains the code for the paper Implicit Weight Uncertainty in Neural Networks (arXiv).

There is a starting point of a reimplementation in Pytorch here.

Abstract

Modern neural networks tend to be overconfident on unseen, noisy or incorrectly labelled data and do not produce meaningful uncertainty measures. Bayesian deep learning aims to address this shortcoming with variational approximations (such as Bayes by Backprop or Multiplicative Normalising Flows). However, current approaches have limitations regarding flexibility and scalability. We introduce Bayes by Hypernet (BbH), a new method of variational approximation that interprets hypernetworks as implicit distributions. It naturally uses neural networks to model arbitrarily complex distributions and scales to modern deep learning architectures. In our experiments, we demonstrate that our method achieves competitive accuracies and predictive uncertainties on MNIST and a CIFAR5 task, while being the most robust against adversarial attacks.

Usage

Following libraries were used for development:

future==0.16.0
jupyter==1.0.0
matplotlib==2.2.2
notebook==5.0.0
numpy==1.14.3
observations==0.1.4
pandas==0.19.2
scikit-learn==0.19.1
scipy==1.1.0
seaborn==0.8.1
tensorflow-gpu==1.7.0
tqdm==4.19.5

Structure

toy_data.ipynb contains the code for the toy regression. The other files contain the code for the mnist and cifar experiments. run_* just calls the experiments. base_layers and layers implement easy to use layers for different VI methods. networks holds the models and the actual training and evaluation is in experiments and utils.

Contact

For discussion, suggestions or questions don't hesitate to contact n.pawlowski16@imperial.ac.uk .

Commands to run experiments:

MNIST:

python run_bbh_exp.py -e 100 -p /vol/biomedic2/np716/bbh_uai/mnist/bbh/ -x layer_a_prior1_fullkernel_noise8 --layer_wise_gen --noise_shape 8 -a --prior_scale 1. --full_kernel -c 0
python run_bbh_exp.py -e 100 -p /vol/biomedic2/np716/bbh_uai/mnist/bbh/ -x layer_a_prior1_fullkernel_noise1 --layer_wise_gen --noise_shape 1 -a --prior_scale 1. --full_kernel -c 0
python run_bbh_exp.py -e 100 -p /vol/biomedic2/np716/bbh_uai/mnist/bbh/ -x layer_a_prior1_fullkernel_noise64 --layer_wise_gen --noise_shape 64 -a --prior_scale 1. --full_kernel -c 0

python run_bbh_exp.py -e 100 -p /vol/biomedic2/np716/bbh_uai/mnist/bbh/ -x layer_a_prior1_fullkernel_indnoise8 --independent_noise --layer_wise_gen --noise_shape 8 -a --prior_scale 1. --full_kernel -c 0
python run_bbh_exp.py -e 100 -p /vol/biomedic2/np716/bbh_uai/mnist/bbh/ -x layer_a_prior1_fullkernel_indnoise1 --independent_noise --layer_wise_gen --noise_shape 1 -a --prior_scale 1. --full_kernel -c 0
python run_bbh_exp.py -e 100 -p /vol/biomedic2/np716/bbh_uai/mnist/bbh/ -x layer_a_prior1_fullkernel_indnoise64 --independent_noise --layer_wise_gen --noise_shape 64 -a --prior_scale 1. --full_kernel -c 0

python run_dropout_exp.py -e 100 -p /vol/biomedic2/np716/bbh_uai/mnist/dropout/ -x dropout_standard -c 0
python run_map_exp.py -e 100 -p /vol/biomedic2/np716/bbh_uai/mnist/map/ -x map_standard -c 0
python run_ensemble_exp.py -e 100 -p /vol/biomedic2/np716/bbh_uai/mnist/ensemble/ -x ensemble_standard -c 1
python run_mnf_exp.py -e 100 -p /vol/biomedic2/np716/bbh_uai/mnist/mnf/ -x mnf_a -a -c 3
python run_bbb_exp.py -e 100 -p /vol/biomedic2/np716/bbh_uai/mnist/bbb/ -x prior_1 --prior_scale 1. -c 4

CIFAR: