Variational Gaussian Dropout is not Bayesian

Abstract

Gaussian multiplicative noise is commonly used as a stochastic regularisation technique in training of deterministic neural networks. A recent paper reinterpreted the technique as a specific algorithm for approximate inference in Bayesian neural networks; several extensions ensued. We show that the log-uniform prior used in all the above publications does not generally induce a proper posterior, and thus Bayesian inference in such models is ill-posed. Independent of the log-uniform prior, the correlated weight noise approximation has further issues leading to either infinite objective or high risk of overfitting. The above implies that the reported sparsity of obtained solutions cannot be explained by Bayesian or the related minimum description length arguments. We thus study the objective from a non-Bayesian perspective, provide its previously unknown analytical form which allows exact gradient evaluation, and show that the later proposed additive reparametrisation introduces minima not present in the original multiplicative parametrisation. Implications and future research directions are discussed.

References in corpus (4)

• Bayesian Compression for Deep Learning
• From $\epsilon$-entropy to KL-entropy: Analysis of minimum information complexity density estimation
• Structured Bayesian Pruning via Log-Normal Multiplicative Noise
• Sparse recovery in convex hulls via entropy penalization

Cited by in corpus (12)

• Learnable Bernoulli Dropout for Bayesian Deep Learning
• Sampling-Free Variational Inference of Bayesian Neural Networks by Variance Backpropagation
• Variance Networks: When Expectation Does Not Meet Your Expectations
• Hypermodels for Exploration
• Semantics Preserving Adversarial Learning
• Epistemic Neural Networks
• On the Effects of Quantisation on Model Uncertainty in Bayesian Neural Networks
• Alpha-Divergences in Variational Dropout
• Analysing Dropout and Compounding Errors in Neural Language Models
• Variational Bayesian Dropout with a Hierarchical Prior
• Practical Bayesian Learning of Neural Networks via Adaptive Optimisation Methods
• Estimating the Uncertainty of Neural Network Forecasts for Influenza Prevalence Using Web Search Activity