Uncertainty quantification with deep learning

This short project, in collaboration with the statistcal machine learning group at NASA Jet Propulsion Laboratory, was focused on developing deep learning models for water quality inversion with quantifiable and scalable uncertainties. Understanding and quantifying uncertainty is a crucial aspect of scientific research and predictive modeling. It provides a measure of the confidence we can place in our predictions. However, traditional deep learning models often overlook the estimation of uncertainty, leading to over-confident predictions. Uncertainty quantification in deep learning is critical to acknowledge the inherent limitations of our models and datasets and to provide more robust and reliable predictions. Two effective techniques for quantifying uncertainty in deep learning are Deep Ensembles and Mixture Density Networks (MDNs).

Deep Ensembles is a method that involves training multiple neural networks (an ensemble) on the same task, each initialized with different weights. The variation in predictions across the ensemble provides an estimate of the uncertainty. This technique has been shown to improve the robustness of the predictions and provide a measure of uncertainty without needing any significant changes to the network architecture. On the other hand, Mixture Density Networks explicitly model the uncertainty in the data by predicting a distribution of possible outputs for each input. MDNs combine a conventional neural network with a mixture model (such as a Gaussian mixture model) to represent the output probability distribution. This approach is particularly useful when the relationship between inputs and outputs is complex and multimodal. Both Deep Ensembles and MDNs offer ways to incorporate uncertainty quantification into deep learning, which is a significant step toward building more reliable and trustworthy AI systems. By acknowledging and quantifying uncertainty, we can develop more robust predictive models that better represent the complexities and variabilities in our data.

These techniques will be incorporated into the operational SWIPE ecosystem for scalable uncertainty quantification.