BayesDLL: Bayesian Deep Learning Library

By Minyoung Kim et al

Published on Sept. 22, 2023

Read the original document by opening this link in a new tab.

Table of Contents

Bayesian Neural Networks: Overview
Approximate Inference Algorithms
Variational Inference
MC-Dropout
SG-MCMC (SGLD)
Laplace Approximation
Uncertainty Quantification
Error Calibration

Summary

BayesDLL is a new Bayesian neural network library for PyTorch designed for large-scale deep networks. It implements various approximate Bayesian inference algorithms such as variational inference, MC-dropout, stochastic-gradient MCMC, and Laplace approximation. The library stands out for its ability to handle very large-scale deep networks including Vision Transformers without requiring extensive code modifications from users. Additionally, it allows pre-trained model weights to serve as a prior mean, particularly useful for Bayesian inference with large-scale foundation models like ViTs. The library is publicly available on GitHub.