Quantifying uncertainty

[agitter]

One of the attractive features of this active learning pipeline is that is explicitly models uncertainty. That has been a shortcoming of traditional virtual screening, especially given the single replicate experimental data that does not allow us to estimate measurement error.

There are many approaches for modeling uncertainty in neural networks or approximating Bayesian models. We should explore some of these to see if they make neural networks more appealing than random forest or other models for this task. Some examples:

• Dropout as a Bayesian Approximation: Representing Model Uncertainty in Deep Learning
• Uncertainty in Neural Networks: Bayesian Ensembling
• Reliable Uncertainty Estimates in Deep Neural Networks using Noise Contrastive Priors

There are arbitrary examples, not necessarily what we should read first or pursue.