Let’s be honest, in recent years, you might have heard of Artificial Intelligence (AI) a million times in your day-to-day life—but have you ever stopped to wonder, how does AI learn to do all the things it can? From writing essays to driving cars and even powering robots, AI has gone from a distant dream to a constant companion.

In fact, many assumes that AI is created with magic that could automatically build itself from scratch and decide on the go. However, the secret of AI is that it is just really good at learning. No magic is involved, just pure machines processing information, recognizing patterns and using these to constantly improve over time.

But what exactly powers all this intelligence? Lend me your attention as I break down three most foundational approaches of how AI learn, mainly supervised, unsupervised, and reinforced learning. Each has its unique way of learning, solving problems, and scenarios where they shine the most.

What even is learning for AI?

It’s called Machine Learning. Think of machine learning as the mastermind behind the greatness of AI. It acts like a brain that learns from acquiring knowledge from data and feedback, then using this knowledge to perform tasks or make predictions. Not much of a difference between us and AI, no? However, the advantage of machine learning over us is that it could learn a large amount of knowledge at a much shorter time compared to us #Overpowered.

Now that we’re aware of what machine learning is, let’s discuss what are the three most basic learning methods in AI.

Supervised Learning: The Teacher’s Pet

You are a teacher teaching a student. As a teacher you give your student guidance with lots of examples with the correct answers for your lecture. The student takes everything and learns through the patterns to understand. 

This is Supervised Learning, the AI model is trained with a labeled dataset, meaning that every input comes paired with the correct output. The model will learn the relationship between the inputs and corresponding labels of the dataset so that when it sees unseen data, it could accurately predict the outcomes.

To better understand how this model learns, check out this diagram:

• In the diagram we have our raw input data (labeled dataset) where there are three classes [bird, cat, dog].
• The model will take this labeled dataset, explore, clean, and split it for the training dataset and testing dataset.
• During training, the model will take into account the relationship between the data and its labels, spotting patterns and features that will help it with predicting who is who later on.
• After training, the performance of the model is validated using the unseen data (testing dataset) using metrics like accuracy, precision, and f1 scores.

This type of learning shines the most when faced with problems like:

• Classification: Wherein the model assigns data points to categories (most common project you could do are spam vs non-spam emails, iris flower classifications and disease diagnosis)
• Regression: Predicting continuous values like forecasting house prices, sales, and salary.

At this point, you may already have an idea what is the difference between supervised and unsupervised learning. Unlike supervised learning who acts like a teacher’s pet, unsupervised learning acts all alone. Let’s further our understanding of it:

Unsupervised Learning: Little Adventurer

Unsupervised learning is said to be an adventurer because it ventures the world of unlabeled data. If supervised learning has help with the labels, unsupervised must understand and uncover the underlying structure and pattern of the dataset on its own. This type of learning shines brightly when we don’t have predefined desired outputs but still want to analyze complex datasets. 

Faced with unlabeled data, unsupervised learning AI models gather information and relation based on similarities and rules that associate the data to each other. This means clustering, associating, or even reducing the dimensionality of the dataset. 

Just as mentioned before, supervised learning shines the most in these types of problems:

• Clustering: the model is grouping similar data points.
• Association: the model tries to understand underlying rules that describe the relationship between each data.
• Dimensionality: Simplifying data while preserving the important details.
• Anomaly Detection: This learning is great in spotting rare or suspicious data points.

What separates unsupervised to supervised learning is just the way the raw input data is presented, simple and easy to understand. However, we still have one more learning approach that sounds massively different from the two priors.

Reinforcement Learning: Human-Like Trial and Error Learner

You’re reviewing for an upcoming exam. If you did well, you get yourself a treat however if you failed to do so, you penalize yourself with no treats. Now, you repeat this process every upcoming exam trying to do your best every time so you can treat yourself. Congratulations! You just got yourself into reinforcement learning.

How reinforcement learning learns is completely different from supervised and unsupervised learning because this type is greatly inspired by how humans learn.

• wherein the model interacts with its environment
• receives feedback in form of rewards or penalties
• then use this information to maximize its next action
• This is also the learning type that we’re most familiar with when AI is mentioned, since this involves an agent that learns and decides.

Reinforcement Learning relies heavily on the training of its agent where the following plays an important role. Think of a robot that tries to solve a maze filled with traps:

That’s a lot, but reinforcement learning is one of the types of learning that AI could do that is so close to being human-like. Reinforcement Learning is all about circling between learning from consequences and striving for continuous improvement.

Supervised vs Unsupervised vs Reinforcement Learning in a nutshell:

At the end of the day, Artificial Intelligence is just one of the tools we can utilize to make our lives a little bit easier. We can either use it to its full potential or let ourselves fear it. However, since we learned that AI is not magic, just understanding the foundation of how these intelligent tools works, not only demystifies the workings of AI, but also opens up the opportunity to look beyond and see it for the powerful tool that it is for solving complex real-life problems.

References:

1. Supervised Machine Learning | GeeksforGeeks
2. What is Unsupervised Learning? | GeeksforGeeks
3. What is reinforcement learning? | IBM