Scientific Applications of Machine Learning Algorithms

Introduction

Machine learning (ML) has а huge impact on modern science. Its impact can even be compared to the impact that mathematics has had in a more traditional scientific context. Scientists from diverse fields either use ML methods directly or adjust them to be better suited for posed problematic. Scientific fields interested in ML are numerous and they include physics, space sciences, chemistry, biology, economy, medicine, etc. The presented overview is non-exhaustive and represents only a sample of all relevant applications of ML in science, organized by categories. The following text will first provide a more detailed insight of recent ML applications in four scientific (sub)disciplines, namely:

• 1.

  computer vision,

• 2.

  bioinformatics and computational biology,

• 3.

  communication networks,

• 4.

  astronomy and astrophysics.

After this, four specific scientific applications will be discussed in much more detail. More precisely, the chapter will present the application of each of the four selected ML algorithms:

• 1.

  neural network algorithm for image classification,

• 2.

  tree-like algorithm for protein structural class prediction,

• 3.

  rule-induction algorithm for intrusion detection,

• 4.

  neighborhood algorithm for the sample selection bias problem in astronomy.

Applications In Computer Vision

Computer vision is one of the busiest areas of artificial intelligence, within which scientists of various backgrounds collaboratively solve complex tasks such as object detection, object tracking, semantic segmentation, etc. Autonomous vehicles represent a technology that is the main driver of new scientific results in computer vision. The reader is referred to (Janai et al., 2020) for a deeper understanding of problems, data sets, and state-of-the-art methods in computer vision for autonomous vehicles. Since there are hundreds of successful ML methods for computer vision, we will mention only some more recent and more relevant ones.

The object detection problem consists in locating and classifying objects visible in natural images. It is similar to the problem of image classification that will be reviewed in the following sections. But unlike image classification, where the goal is to assign a class to an input image, object detection also needs to localize one or more objects on a single image with bounding boxes, and then further classify each one. Solutions to this problem are usually modeled as software systems that are tightly coupled with underlying hardware, such as video, thermal infrared cameras, or laser scanners. It matters to have better hardware because the software system can be fed by more useful information, i.e., more features and/or better feature measurements. There is a vast number of methods for completing the object detection task by employing deep neural networks—see the recent review by Zhao et al. (2019).