CS Forum: Tommi Kärkkäinen "Distance-based Machine Learning Methods for Nanoclusters"

Distance-based Machine Learning Methods for Nanoclusters

Tommi Kärkkäinen
Professor, Faculty of Information Technology
University of Jyväskylä

Abstract:

One could try to circumvent the complex and time-consuming simulations related to nanostructures by using supervised machine learning methods. Different structures then need some sort of feature representation (descriptor) of fixed size, and recently many such derivations have been proposed and analyzed. Distance or dissimilarity has a central role in unsupervised learning, e.g., in clustering algorithms. With supervised scenario, the use of distance-based methods can be traced back to linear Radial Basis Function Networks (RBFN) from 1980s. I will introduce two rather recently proposed distance-based random-basis methods: Minimal Learning Machine (MLM) and Extreme Minimal Learning Machine (EMLM). Both of these methods base the construction of a nonlinear regression model solely on distances within the training data. In the talk, these methods and their applications for nanostructures will be depicted. Interestingly, both building blocks of the talk - desciptors for nanoclusters and MLM - are directly linked to the research in the Aalto University.

Prof. Tommi Kärkkäinen will act also as an opponent in Alexander Grigorevskiy's defense "Advances in Randomly-Weighted Neural Networks and Temporal Gaussian Processes" on Friday 20 September at noon in lecture hall T2, CS building.