Introduction

Overview

Teaching: 15 min
Exercises: 10 min

Questions

• What is machine learning?

• Where is machine learning useful?

Objectives

• Learners should have a basic working understanding of machine learning.

• Learners should have a basic understanding of the problems where machine learning may be applied.

How do we make sense of large data sets? How do we distinguish different groups of objects?

In our modern world, there are many problems where one might want to distinguish different types of objects into categories or groups. Consider the following real-life examples:

• An e-mail provider might wish to separate spam e-mails from non-spam e-mails so that its customers aren’t inundated with messages they don’t want to receive.
• In the postal system, thousdands of letters, postcards, etc. are received every day and these need to be sorted by location and address. Sorting efficiently is paramount, but one complication is that different people have different handwriting. For example, the post needs to make sure that an “a” is always recognized as an “a”, so that addresses are parsed correctly and letters get sent to the right addresses.
• Maybe you’re an amateur photographer (like me!) and have photographed a lot of cats and dogs. You now want a way to distinguish your cat photos from your dog photos, perhaps because you’d like to print out some cat photos for your cat-loving friend.
• Maybe you’re a scientist who has recorded audio sounds from a forest for an entire spring and summer. Now you wish to differentiate between types of bird calls in order to estimate how many different species there are.

All these are problems could be done by hand because humans are extremely good at classification. For example, you could sit down and sift through all of your photos and decide whether there’s a cat or a dog in each one. You’ll probably also be able to note some specific differences between both (maybe the shape of the head, or the ears, or the presence of whiskers) that help you make that distinction. Once you’ve seen a picture of a dog or a cat, you’ll probably also be able to abstract from that picture to a new one. If on that new picture, the cat or dog is shown from a different angle, chances are you’ll still be able to extrapolate and tell me what kind of animal it is.

However, you might not want to sit down and sift through all of your photos by hand. If you’ve got many hundreds or thousands of photos (perhpas you’re a photographer who specializes in pets), that might become very tedious, time-consuming and boring. Sorting might also not scale very well. For example, each day the US Postal Service processes something like 187 million pieces of mail! Imagine if each of those addresses had to be read and typed in by a postal worker by hand. Wouldn’t it be cool if we could make a computer help us classify those images of cats and dogs, or help process the mail?

This is where machine learning comes in. Machine learning is a subfield of computer science concerned with methods and ways that we can teach computers to learn patterns from data. Often, the goal is classification: distinguishing photos of cats from those of dogs, looking for similar songs to the one you’ve just been listening to, or even providing medical diagnoses. In these types of problems, we are trying to predict whether a given example belongs to one of several discrete classes. In this tutorial, we will focus on problems with this kind of structure. It’s worth nothing, however, that machine learning can also solve problems where the output variable is continuous, where we are trying to perhaps predict the temperature in Seattle tomorrow from past weather data. This type of problem is called regression.

Challenge

Take a moment to think about where you might have encountered classification problems in your daily life. Can you think of places where your life is affected by computers making decisions on your behalf?

The Process of Machine Learning & This Tutorial

As mentioned above, in this tutorial, we are going to focus on the problem of predicting the discrete class of an object–cat versus dog, spam versus not spam–from a set of data points, e.g. an image or the collection of words in an e-mail. In machine learning, we do that by building a model, i.e. a mathematical function that maps between the data points (photo) and the class (cat). This requires a training data set, i.e. a data set for which we know what the correct class for each example is. In many cases, providing raw data (e.g. the red/green/blue values for each pixel in an image) is at best an inefficient way to do machine learning. Instead, we often extract features, descriptive summaries of the data designed to be as different as possible for our different classes. For example, if you wanted to classify pictures of apples versus pictures of pears, the total number of red pixels in the image might be a descriptive feature. Machine learning models are trained by providing the algorithm with the training data and the class labels it maps to, and adjusting the model parameters (numbers are free to vary and adjust the functional form of the model) until you find the parameters that best fit your data (there are computer algorithms that can do that, too). Once you have built the model, you need to evaluate its performance: How will you know if your model does a good job? How do you know whether it is reliable? Does it manage to classify all or most of the examples you showed it? Does it preferentially classify some classes well, but not others?

It is important to be aware that building a machine learning model involves a myriad of decisions big and small that are part of the process. For example, different machine learning algorithms provide different types of functions to build the desired mapping and are therefore useful for different kinds of problems. Some algorithms return the predicted class of the example data you showed it. Others provide probabilities for all the available classes, and the user must decide how to interpret these probabilities. In order to build a useful machine learning model, you need to understand the structure of the problem you are trying to solve, and the properties of the available training data. How many classes are there? Are you sure that your training data captures all the available classes? How many different examples do you have for each class? In addition, each algorithm usually has different options one can adjust, and the model builder must decide whether to set these based on the properties of the problem they’re trying to solve or include these options in the training process as parameters.

Many of the considerations and questions are ethical in nature. Classification is almost never the end goal, but machine learning is very often used to support human decision making. For example, a spam detection algorithm might decide whether to show the user a specific e-mail, or auto-delete it. But these decisions have consequences for the people that are affected by the algorithm, and so when building a system, one must think about how the decisions made based on an algorithm will affect human lives.

This tutorial aims to give an introductory overview of the process of machine learning. Rather than being comprehensive and mathematically rigorous, it aims to highlight the key steps that are part of building a successful model, and point out the pitfalls that generally occur when building these models. The ethical questions related to building machine learning models–and the ways in which these things can to very wrong–take a central role in this tutorial, reflecting their important position at the heart of the model building process.

This tutorial owes a lot to previous machine learning tutorials at [Astro Hack Week][ahw], especially Iain Murray’s and Gilles Louppe’s tutorials and Adrian Price-Whelan and David W Hogg’s tutorial. It is also a follow-up on Gwen Eadie’s paper on teaching Bayesian statistics with M&Ms. Because what better way to learn statistics and machine learning through chocolate?

Challenge

Can you think of situations in your own life where you think implementing a machine learning model might be useful to support your decision making process?

Key Points

• Machine learning is the study of algorithms to learn patterns from data.

• Machine learning is ubiquitous in the modern world.