deepbayes is a package that aims to make training Bayesian neural networks (BNNs) simple.

Implemented on top of TensorFlow 2.0 and working with tf.keras as a primary interface, one can simply declare a keras (or TF) model and pass it to the desired deepbayes optimizer and it will compile the model into an object that is ready to perform approximate inference.

We also support a variety of adversarial robustness analysis methods for Bayesian Neural Networks.

deepbayes is actively under development, and location and names of packages may change.

Current version 0.1.0 (Updated 27 May 2022)

pip3 install deepbayes

In addition to producing posterior distributions for BNNs, deepbayes also supports some basic analysis including:

• Adversarial attacks for BNNs (FGSM, PGD, CW)
• Statistical Guarantees for the Robustness of Bayesian Neural Networks - https://arxiv.org/pdf/1903.01980.pdf
• Probabilistic Safety for Bayesian Neural Networks - https://arxiv.org/pdf/2004.10281.pdf
• Adversarial Robustness of Bayesian Neural Networks - https://bit.ly/3lLqENo
• Bayesian Inference with Certifiable Adversarial Robustness - https://arxiv.org/abs/2102.05289

• Uncertainty Quantification with Statistical Guarantees in End-to-End Autonomous Driving Control - https://arxiv.org/pdf/1909.09884.pdf
• Robustness of Bayesian Neural Networks to Gradient-Based Attacks - https://arxiv.org/abs/2002.04359
• Gradient-Free Adversarial Attacks for Bayesian Neural Networks - https://arxiv.org/pdf/2012.12640.pdf

Implemented based on description in https://arxiv.org/abs/1505.05424

Implemented based on description in https://arxiv.org/abs/1206.1901

Implemented based on description in https://arxiv.org/abs/1712.02390

Implemented based on description in https://arxiv.org/abs/1806.04854 and https://arxiv.org/abs/2002.10060

Implemented based on description in https://arxiv.org/pdf/1902.02476.pdf

Original paper by Robbins and Monro

Implimentation based on description in https://arxiv.org/abs/1412.6980

tensorflow, tensorflow-probability, numpy, tqdm, statsmodels,