Research

Machine Learning

Topic Overview

When making a computer calculate something, we write a procedure manual (program) that shows what to perform in what order, and then execute it. Usually, when solving problems that can be defined using mathematical formulas, it is clear what kind of processing should be performed, and programs are often easy to write. However, for other tasks, even if they are easy for humans, for example, transcribing what is spoken from the recording of a conversation, or describing the contents of a photograph, writing a program is very difficult. Because the data handled by the computer is an enumeration of audio waveform information and red/blue/green color information, it is almost impossible to mathematically describe the process of outputting what is being said or to determine what image is being used.

That’s where machine learning comes in. In machine learning, patterns are extracted from a lot of known data, and knowledge is acquired so that even unknown data can be correctly judged based on the extracted patterns. There are various methods for machine learning, but in recent years, methods such as1 a deep learning that uses a mechanism similar to the signal transmission of neurons in the brain have made it possible to recognize images with high accuracy. The range in which machine learning can provide sufficiently reliable accuracy is expanding rapidly in other fields as well.

Development at ICReDD

ICReDD aims to discover new and truly useful chemical reactions from data obtained from experimental chemistry and a vast amount of data from computational chemistry. To this end, we are developing machine learning methods, aiming at predicting the reactivity of unknown reactant combinations and the physical properties of products2.

We are also working on the development of algorithms that can be applied to efficient scheduling through machine learning in order to reduce the time and economic costs of computational chemistry and chemical experiments that must be repeated to discover new chemical reactions3.

References
  1. ImageNet Large Scale Visual Recognition Challenge http://image-net.org/challenges/LSVRC/
  2. Alexandre Varnek and Igor Baskin, “Machine Learning Methods for Property Prediction in Chemoinformatics: Quo Vadis?”, Journal of chemical information and modeling, 2012.
  3. K. Tabata, A. Nakamura, J. Honda and T. Komatsuzaki, “A Bad Arm Existence Checking Problem”, Machine Learning, accepted for publication.