QED has been involved with the development of AI for many industrial projects requiring classification and prediction, including:
- optical character recognition on real-world medical data
- croplands and building footprints from satellite imagery
- spectral prediction of particular nutrients in soils
Our team’s high fluency in both computer science and statistics allows us to work from both cultures of statistical modeling perceived by Leo Breiman. We can combine modern approaches — such as deep learning, random forests, and SVMs — with more classical statistical analysis such as ANOVA, random-effect models, Bayesian Additive Regression Tree (BART) models, and spatial-temporal analysis. Many of our current AI applications are related to computer vision, described in more detail here.
Lastly, our philosophy about AI is that it is actually the last resort, when more classical approaches have failed. This is often the case with natural systems in health and agriculture, where the complexity of the subject matter often escapes humans’ modeling ability. However, much opportunity for mathematically-driven optimization, based on formal proof, still exists in many man-made systems. Some examples include industrial/chemical engineering plants, logistics, and satellite and telecommunications systems.