Skip to content
Main menu hidden.
Published: 30 Jun, 2021

Old lessons in chemometry can improve modern AI - and vice versa

NEWS With the increased use of automatic systems, it is important that we understand the limitations in all parts of the chain. Rickard Sjögren has explored the interface between chemometry and AI to find methods that can improve the conditions for understanding. He defended his dissertation on Friday 28 May at Umeå University.

Text: Ingrid Söderbergh

Artificial intelligence, AI, is on the lips of many with great promises about what it can contribute to society. At present, however, most of the AI ​​used is quite narrow and consists of what is called machine learning. Thanks to the fact that machine learning has mathematical methods for developing computer programmes that learn to perceive patterns based on examples, rather than being programmed by a human, it is extremely useful in many parts of society and industry.

Because machine learning is so useful, today large amounts of data are collected to be used as examples. In chemistry, for example, the field of chemometry aims to develop methods for extracting knowledge from complex chemical data to optimize or investigate chemical processes. This is done, among other things, with the help of machine learning, but also multivariate statistics.

“Chemometric methods have been used in practice since the 80s, which has given solid experience of peculiarities of data from real processes and how to ensure that one's methods do what they think they do. This practical approach has led to chemometric methods being used today in, among other things, quality assurance in the tightly regulated pharmaceutical industry” says Rickard Sjögren, doctoral student at the Department of Chemistry at Umeå University.

In his dissertation, Rickard Sjögren has explored the borderland between chemometry and machine learning. There are many lessons in chemometry that can be used in typical machine learning applications. At the same time, machine learning is a much larger field with an enormous accumulated knowledge that can help chemometry move forward.

“Despite the fact that both fields have a lot in common, cross-pollination between them is quite limited, which is why I show in my dissertation examples of how we can improve the yield.”

For example, Rickard Sjögren has been inspired by chemometry to develop a method that allows a certain type of machine learning methods, common in more advanced AI applications, to detect when they are unsure of what they understand by data.

“These methods are used, among other things, for computer vision that interprets the environment for decision-making systems in self-driving cars. Then it is extremely important that the systems can check when they are unsure of what they are seeing in front of the car to avoid traffic accidents. There are already methods to solve this problem, but the method I have developed is both reliable and computationally efficient compared to the alternatives.”

Creating really good machine learning systems requires huge datasets, but ensuring the quality of huge datasets is not always easy. In another study, Rickard Sjögren participated in creating a large dataset for microscopic image analysis, which has not previously been done on such a large scale.

“Due to the large number of microscopic images and their complexity and the fact that many people were involved, we had to carefully plan the work to ensure quality. Among other things, we used methods for experimental planning that are common in chemometry to reduce the risk of systematic errors creeping in.”

Through his dissertation, Rickard Sjögren hopes to provide inspiration for how two fields that have a fairly limited exchange with each other can benefit in more ways than one might have thought before.

“It is inevitable that there is a lot to be gained from the enormous pace of development we see in machine learning. But at the same time, there are many lessons from other smaller fields that have long focused on practical use.”

Rickard Sjögren was born and raised in Vårgårda in Västergötland but has lived in Umeå for ten years. He has a master's degree in biotechnology with a focus on bioinformatics from Umeå University. Since 2018, in parallel with his doctoral studies, he has worked full time as a researcher in advanced data analysis at Sartorius, a global supplier to the biopharmaceutical industry.

Read the whole dissertation

Press photo

About the dissertation
On Friday 28 May Rickard Sjögren, Department of Chemistry at Umeå University defended hos thesis titled Synergies between Chemometrics and Machine Learning. Swedish title: Synergier mellan kemometri och maskininlärning. Faculty opponent was professor Ola Spjuth, Department of Pharmaceutical Biosciences, Uppsala University, Sweden.