Tools for uncertianty estimation in Deep Learning models. Use the below command in REPL to install.
] add DeepUncertainty
Neural Networks are usually trained to minimize a loss function that approximates model performance on any given task. The weights (parameters) of the network are estimated using first-order gradient based optimization algorithms that usually result in a point estimates.
We can also take take a Bayesian view point and place a distribution on each weight instead of interpreting it as a point estimate. We can calculate the uncertianty in our estimates of the parameters, since we optimize the distribution parameters insteaf of point estimates directly.