The AI ABCs

The AI ABCs#

This Jupyter book is a high-level introduction to using machine learning (ML) and artificial intelligence (AI) methods in scientific research. Its design has been heavily inspired by Software Carpentry, a program run by the Carpentries. Whereas Software Cerpentry is a high-level introduction to basic programming tools, this course is a high-level introduction to basic ML and AI tools.

Accordingly, this course is intended for readers who have programming experience but who do not necessarilty have strong mathematics and statistics backgrounds. (See the prerequisites page for more information.) It attempts to provide students with intuition about how various ML techniques work, how they can be used, and what their limitations are without delving deeply into theory.

What kind of Artificial Intelligence does this course cover?#

AI can mean many things: large language models (LLMs) that can communicate like humans, reasoning engines that can deductively prove theorems, and algorithms that learn the structure of data, to name a few. The term “AI” is often associated with artificial general intelligence, which refers to an intelligence that excels at all cognitive domains a human typically excels at. Large language models (LLMs), though not generally intelligent in the manner of humans (as of when this course was written), are nonetheless somewhat good at answering questions across a very broad set of domains. Other kinds of AI tools are good at solving only a very narrow set of problems.

This course focuses on a particular subset of AI that is useful for exploring and understanding structured numerical data. The “numerical” in “numerical data” just means that the data are at least partly composed of numbers, usually real numbers. Some machine learning algorithms in this course can be adapted for use with “categorical” data, but these topics are largely out of scope for this course. Many datasets include both numerical and categorical dimensions—for example, one might measure both a child’s reading speed (a numerical variable) and their handedness (a categorical variable)—and this course will occasionally discuss how to handle such variables in numerical algorithms.

The “structured” in “structured numerical data” means that the data have an inherent model of how their pieces are related to each other. For example, if we measure the trajectory of a fly over time by recording its \((x, y, z)\) position coordinates every millisecond for 5 seconds, our data would be structured in multiple ways: the \(x\), \(y\), and \(z\) coordinates are orthogonal axes, and the measurements would be ordered in time. If we choose two arbitrary pieces of the data, say the \(x\) coordinate at time 0 and the \(y\) coordinate at time 1, we can say how they are related to each other. This differs from “unstructured data” in which there is no clear relationship between any two pieces. For example, a corpus of text and a set of JSON files can each contain rich and detailed data, but without knowing something else about the text or files, we can’t say how any two arbitrary words or JSON entries are related to each other. If you are looking for guidance on unstructured data, such as corpuses of text or collections of JSON files, we suggest that you look into Natural Language Processing (NLP) and LLMs.

Table of Contents#

Introduction

Unsupervised Learning

Supervised Learning

PyTorch and Optimization

Neural Networks

About#

Authors#

Noah C. Benson <nben@uw.edu> (Maintainer)
eScience Institute
University of Washington

Contents