Pyesian 0.0.1

A framework for Bayesian Neural Nets

Introduction

Many successful pieces of software today involve machine learning, which is why many tools and libraries were created to make implementing machine learning models easier. The best examples are Tensorflow, PyTorch and Keras. We want good software to accelerate progress in ML. One area that has a lot of promise to accelerate progress is Bayesian deep learning.

Deterministic NNs seem to have many failure modes including discriminatory bias, poor uncertainty, and a lack of robustness. An important source of these is that the choice of the weights $\Omega$ of our neural network is often highly discriminating. $\Omega$ is often one local minimum among many others and does not encompass all the 'knowledge' needed to solve tasks. For instance, in a classification problem the choice of a particular $\Omega$ does not encode the uncertainty in the predictions being made. As long as our input point $x$ is 'far enough' from the decision boundary, the NN will predict class A with high probability, even if no data points exist around our input $x$. This can be seen on the illustrative plot below.

An alternative that can potentially remedy the aforementioned failure modes is Bayesian learning. Bayesian machine learning involves inferring a posterior distribution $p(\Omega|D)$ over model parameters via Bayes rule given the dataset $D$: $$p(\Omega|D) = \frac{p(D|\Omega)p(\Omega)}{\int p(D|\Omega)p(\Omega)d\Omega}$$ This rule has a prior $p(\Omega)$ which expresses prior knowledge and has been shown to have nice trustworthiness properties. Though extremely valuable to have, one cannot compute the exact posterior. Thus, approximate inference methods have been proposed. These diverse approaches to Bayesian inference, especially for deep learning, have not been united in a way that would enable Bayesian ML to have the same acceleration we have observed for deterministic ML.

There exists many inference methods to approximate posterior distribution (HMC, SGLD, SWAG, BBB...), each with its pros and cons (runtime, accuracy of approximation...).

Our library builds on top of Tensorflow Keras © to make the training of Bayesian Neural Networks simple. It groups together a wide range of inference methods in one easy-to-use framework to enable Bayes in ordinary NNs with minimum efforts on the user-side as well as many other features listed below.

Inference Methods implemented

• SWAG
• Bayes by Backprop (BBB)
• SGLD
• HMC
• SGD
• SGLD
• ADAM
• VADAM
• BSAM

External libraries integrated

• Support to all Kersa models architectures
• Support to all Tensorflow probability distributions
• Support to Weights and Biases
• Support to OpenAi Gym RL envirenment
• Tensorflow datasets easy loading

References

• Tensorflow: https://www.tensorflow.org/
• Tensorflow probability: https://www.tensorflow.org/probability
• Tensorflow datasets: https://www.tensorflow.org/datasets
• Weights and Biases - The AI developer platform : https://wandb.ai/site
• Improving PILCO with Bayesian Neural Network Dynamics Models: https://www.cs.ox.ac.uk/people/yarin.gal/website/PDFs/DeepPILCO.pdf
• MCMC Using Hamiltonian Dynamics: https://www.mcmchandbook.net/HandbookChapter5.pdf
• Bayesian Learning via Stochastic Gradient Langevin Dynamics: https://www.stats.ox.ac.uk/ teh/research/compstat-s/WelTeh2011a.pdf
• A simple baseline for Bayesian uncertianty in deep learning https://arxiv.org/pdf/1902.02476.pdf
• Weight Uncertainty in Neural Networks https://arxiv.org/pdf/1505.05424.pdf
• Simple and Scalable Predictive Uncertainty Estimation using Deep Ensemble https://arxiv.org/pdf/1612.01474.pdf
• Fast and Scalable Bayesian Deep Learning by Weight-Perturbation in Adam: https://arxiv.org/pdf/1806.04854.pdf
• Sam as an optimal relaxation of Bayes: https://arxiv.org/pdf/2210.01620.pdf