HoCo Hour of Code

I'm sure you've heard of artificial intelligence over the past couple of months. With the rise of products like Chat-GPT, Snapchat AI, and Bing's new AI search assistant, it's kind of hard not to. In fact, chances are you’ve probably played around with some of these bots. However, these trendy new tools are only a small subset of AI. Artificial intelligence (abbreviated as AI) is any sort of program that tries to mimic human intelligence and capabilities in order to solve a problem.

These artificial intelligence systems are built upon ML (machine learning) algorithms. Machine learning algorithms are trained on data which they use to apply what they have “learned” to make decisions. A wide spectrum of programs can be considered machine learning programs, from the linear regression function on your calculator to complex medical diagnosis programs.

Now, you might ask, “When would we even need machine learning algorithms when humans could perform the same functions?” AI has two significant advantages over humans:

It can do things faster than humans
It can do things more accurately than humans

The copious amount of data required to solve certain problems, along with the margin for human error, make them a perfect field to apply machine learning applications to.

Well, how does a ML model work then? At a high level, a machine learning model is basically just a regular old function. In math class, your teacher probably covered a function, defined as y=f(x). Given a value x, the function f(x) would output a value of y. Now imagine that you only know the y-value in certain cases, but want to predict the y-value in cases where you don’t know the y-value. That is where a basic prediction function comes into play. The goal of machine learning algorithms is to find the optimal function f(x) such that our model can properly predict an accurate value y for any given x. This function could be anything from a simple linear function to an extremely complex function.

Let’s say you have a dataset of the accuracy of NFL quarterbacks versus their arm lengths, with arm length being the x-value and accuracy being the y-value. If you wanted to predict the accuracy of a quarterback with an arm length of 3.5 feet, but no such quarterback exists in the NFL, you could use a machine learning model to predict the hypothetical quarterback’s arm length. Let’s walk through the steps for this.

First, you would split the given dataset you have into a training set and a testing set for your model. It is important to split your data so that you have enough data to accurately train the robot and also measure its performance on data that it hasn’t seen before. Then, you would choose between different types of models to find a f(x) that best fits the given trend of accuracies and arm lengths. After using our training set to find the best f(x), we would use the test set to evaluate the function and determine how accurate it is at generalizing the trend.

In the above sample, we only have one X variable (arm length) and one Y variable (accuracy). However, machine learning models applied in scientific and commercial fields can have hundreds or thousands of variables. This is where a distinction between real and ideal data can be made. Ideal data has an easily generalizable trend with strong correlations between variables. Real data is data in the real world that does not always have strong correlations between variables.

Obviously, we would always prefer ideal data to real data. However, most machine learning problems have messy data; you can’t always just use the same model to solve every problem. Different problems require different models and types of learning. Luckily, if trained correctly, a model can generally extract patterns from messy data.

Within machine learning, there are two main ways a model learns: supervised and unsupervised learning. Imagine you are trying to classify whether an email message is spam or not. Supervised learning would give us each input X and label the corresponding output Y as spam or not spam. On the other hand, unsupervised learning would give us each input X, but would not label whether the corresponding output Y is spam or not. You can think of supervised learning as providing a strict set of rules for the computer to follow while unsupervised learning gives the computer more liberty to pick up subtle patterns on its own.

TThere is also a third type of machine learning called reinforcement learning. In reinforcement learning, a ML algorithm learns through trial and error. The algorithm gets rewarded and penalized for particular actions. Reinforcement learning algorithms think several steps ahead, as they always attempt to make the optimal move. Reinforcement learning is used in specialized cases such as video games with AI opponents, travel planning algorithms, and budget optimization algorithms.

In order to find the optimal function f(x), we need to find the parameters to use for that function. You can think of parameters as things that turn an input into an output. For example, the classic equation of a line, y = mx + b, has the parameters m and b. A machine learning algorithm typically goes through the data and makes small tweaks to its parameters until it can find the best possible function. Different learning algorithms and model types do this in different ways.

There are two main tasks that can be accomplished with ML: regression, and classification. Regression involves making quantitative predictions on continuous data. Classification involves putting data into qualitative categories.

Machine learning has a variety of applications in our day-to-day lives from social networks to personalized medicine to recommendation algorithms and navigation. Understanding the foundations of ML and how to use it will become an increasingly important skill in years to come. In future articles, we will focus more on regression and also dive into deep learning, an increasingly popular subset of machine learning.