Embrace modeling uncertainty with machine learning

When we were in school, we probably believed that science and math were primarily about offering definitive answers. But adult scientists and engineers know that’s rarely the case.

Uncertainty is inherent in almost all systems studied and experiments conducted. This may arise from the natural scatter of experimental results, limitations in the precision of measurement methods, errors made while performing measurements, or the need to extrapolate from incomplete data.

Image Credit: Intellegens Limited

We also have to deal with uncertainty in machine learning (ML). Predictions from ML models will be uncertain partly due to the inherent scatter of training data and partly because the models are built from limited data.

It is important to embrace this uncertainty, rather than expecting to eliminate it entirely. It must be understood, quantified and used to make better decisions.

In fact, uncertainty is a positive. Just as a straight line that passes through all the points on a graph reveals how simple a system is to characterize, a model containing negligible uncertainty is unlikely to be very useful or insightful. These are the difficult and complex problems that come with uncertainty, and these are the problems that are important to solve.

ML specialist Intellegens has taken every precaution to ensure that uncertainty is accurately quantified in the predictions of its Alchemite™ algorithm. The algorithm extends conventional methods which, at worst, do not estimate uncertainty at all but assume a normal distribution, or capture only one of many contributions to the overall uncertainty.

Accurate quantification, for example, helps identify the probability of success of an experiment so that the user can target experimental work accordingly.

Intellegens CSO Dr Gareth Conduit presented some exciting new research from his group at the University of Cambridge in a recent webinar. The work done by Dr. Conduit and his team goes a step further in accepting uncertainty.

He focused on the design of concrete, for which the measurements of many parameters associated with the material appear “noisy”.

Rather than simply quantifying the uncertainty, this work allowed the machine learning method to learn from noise in the data, discovering that the uncertainty of a particular physical parameter could be used to help predict the strength of the concrete.

Two new concrete mixes were proposed as a result of this work, subsequently created, and then found to behave as expected, which expanded the properties of commercially available mixes.

If we’re looking for one certainty in all of this, it’s that machine learning can do some interesting things in the world of chemicals and materials. But only if you take the right approach to uncertainty!

This information has been obtained, reviewed and adapted from materials provided by Intellegens Limited.

For more information on this source, please visit Intellegens Limited.

Sherry J. Basler