Probabilistic Machine Learning Series Post 3: Weights Uncertainty with Correlated Noise

Using correlated dropout to quantify the uncertainty of Neural Network forecasts

One of the main drawbacks of the Bayesian approach is that it is not scalable. In 2015, Blundell & al. proposed to quantify the uncertainty in Neural Networks by using dropout. The method, called Bayes by backprop, uses the reparametrization trick to approximate a variational bound to the posterior. In this post, we will experiment with Bayes by backprop by making a few adjustments to the method. …

bidirectional-lstm data science hands-on-tutorials machine machine-learing machine learning noise probabilistic-models series uncertainty