skip to content

Cambridge Service Alliance

At the forefront of service transformation in the digital era
 
High-Quality Prediction Intervals for Deep Learning: A Distribution-Free, Ensembled Approach

February Paper on High-Quality Prediction Intervals for Deep Learning: A Distribution-Free, Ensembled Approach by Tim Pearce, Mohamed Zaki, Alexandra Brintrup, Andy Neely

[Paper]

Deep neural networks (NNs) have caused great excitement due to the step-changes in performance they have delivered in a variety of applications. However, their appeal in industry can be inhibited by an inability to quantify the uncertainty of their predictions. To take a prognostics example, a typical NN might predict that a machine will fail in 60 days. It is unclear from this point prediction whether the machine should be repaired immediately, or whether it can be run for another 59 days. However, if the NN could output a prediction interval (PI) of 45-65 days with 99% probability, timing of a repair could easily be scheduled. In this paper, we develop a method for doing exactly this - the quantification of uncertainty in deep learning using PIs. We derive a method based on the assumption that high-quality PIs should be as narrow as possible, whilst still capturing a given proportion of data. The method is general, applicable to any data-driven task where a continuous value needs to be predicted, and it is important to know the uncertainty of that prediction. Examples include the forecasting of precipitation, energy load, financial metrics, and traffic volume. The method is tested on ten real-world, open-source datasets. The proposed method is shown to outperform current state-of-the-art uncertainty quantification methods, reducing average PI width by around 10%.

Cambridge Service Alliance

Welcome to the Cambridge Service Alliance…

  • A unique global alliance between the University of Cambridge and some of the world’s leading businesses.

  • Help organisations to address the challenges they will face in the next three to five years, through rigorous research, practical tools, insights and education programmes.

  • Learn how other innovative organisations are developing new services through our events

  • Since its inception in 2010 industrial partners have included BAE Systems, Caterpillar, GEA, IBM, Pearson, Zoetis, CEMEX, HCLTech and Bouygues


CSA News

Lastest News.....