Supervised, unsupervised, reinforcement learning... what is AI learning?

For AI to be effective, it needs to learn to grow. There are three main types of learning used to train artificial intelligence. Understanding how AI learning works provides a better understanding of how these computer programs, which are increasingly present in our lives, function. How many types of learning are there, and what exactly are they used for? We explain it all here.

Photo de randa marzouk sur Unsplash

In short

• Supervised learning trains AI with labeled data for tasks such as recognition or prediction.

• Unsupervised learning works without labels and detects patterns or anomalies in data sets.

• Reinforcement learning relies on a system of rewards and punishments to find the best possible action.

• Each method has specific uses: real estate, finance, marketing, video games, sentiment analysis, and even fraud detection.

• The future of AI could combine these approaches into hybrid models that are even more powerful and versatile.

Supervised learning: what is it and what does it mean?

Supervised learning involves training AI with labeled data. For example, let's say we want to create an AI that can automatically recognize vegetables in a supermarket, making it easier for customers to identify and select the right type of vegetable when shopping. To do this, we will compile a large database of images containing several vegetables. These images will then be labeled. This means that they will be described and prepared in advance by humans to tell the AI directly what a zucchini, a pepper, etc., looks like. The artificial intelligence will therefore be trained in a supervised manner, under human supervision.

What can AI trained with supervised learning be used for?

Artificial intelligence built using supervised learning can be used for recognition and analysis tasks, or even prediction. For example, AI in this category can be used for sentiment analysis. By showing the AI what a smile, a sad face, an angry person, etc. look like using labeled data, it is then able to analyze people's facial expressions to determine their state of mind.

We can also mention broader areas such as real estate. For example, by giving AI labeled and accurate data on a property, such as the price it should have for a certain surface area and other characteristics. A real estate agency could then have a property valuation tool on its website for people wishing to sell their home and get a quick and reliable estimate.

Unsupervised learning: what is it and how does it differ from supervised learning?

Unlike supervised learning, unsupervised learning does not involve labeling the data sent to train the AI. In concrete terms, the AI must learn to recognize patterns that may exist in data packets on its own to more effectively detect possible abnormal patterns and structures. It can therefore be said that artificial intelligence trained through unsupervised learning is more intelligent, in the sense that it will be better able to detect anomalies and find patterns that humans would not have thought of. However, it is less effective for conventional detection than AI with supervised learning.

Unsupervised learning, therefore, works by classification and association. AI will seek to classify the data it has, to create broad categories that will then enable it to associate this data in order to create links and groups. It will thus be able to detect patterns and trends in this data. For example, if we are talking about AI used by a newspaper to categorize articles, it will try to associate all articles that discuss ecology to group them.

What can AI trained with unsupervised learning be used for?

We can assume that AI trained with an unsupervised learning model could be used in the field of finance, for example, to detect fraud. By learning classic patterns of transactions, investments, and financial transfers, artificial intelligence would be able to detect anomalies and irregularities more quickly than a human.

More specifically, "Google News uses unsupervised learning to classify articles on the same topic in different online media."* Another example is marketing, where unsupervised learning enables AI to classify and associate certain parameters and purchasing habits to create or update profiles of potential buyers. Artificial intelligence can therefore be used to detect new profiles and new business opportunities.

*IBM, qu'est-ce que l'apprentissage non supervisé ?

Reinforcement learning: what is it?

Reinforcement learning is also a machine learning method, but this time it focuses on finding the best action to achieve the highest performance. In concrete terms, the software is trained to make the best decisions to achieve the best possible results. This type of learning is similar to trial and error. In other words, the AI will try several different approaches to achieve a goal and will analyze which one is the most effective and optimal. This system is based on a system of punishment and reward. The goal is to obtain as many rewards as possible.

What can AI trained with reinforcement learning be used for?

In practical terms, this type of artificial intelligence is ideal for investment. In this sector, where rates and markets fluctuate from day to day, and even from minute to minute, reinforcement learning AI will constantly analyze the most effective method of achieving its objective, such as finding the best rate at which to invest a sum of money.

This learning system, therefore, allows AI to learn from its mistakes in order to improve and seek ever more efficient solutions. This is particularly true of AlphaGo, Google's AI that has beaten the world Go champion several times. By playing against him over and over again and running simulations with a reward/punishment system, the AI has improved to the point where it has become unbeatable.

Another example: video games! It's easy to imagine an AI capable of adapting to the player's behavior as they progress through the game. This is partly the case in the video game Alien Isolation, where the artificial intelligence created for the alien adapts to the player's style of play and seeks to interfere with their progress. The AI used for the alien is divided into two parts: one AI that constantly knows the player's exact position and status on the map, and a second AI that searches for the player in the player's "global" area, reacting to the player's noises and other interactions in the game. While the first AI is never allowed to share the player's position with the second for obvious reasons of cheating, immersion, and gameplay, the second must therefore find them using approximate information. The alien's artificial intelligence also acts using a gauge that determines the player's potential stress level. This gauge uses parameters such as the time spent near the player, the distance between the alien and the player, etc. Once the gauge is too high, the AI will scare the alien away to allow the player to progress through the game and not be in a state of constant tension, which would detract from their gaming experience.*

*Youtube, AI and games ; The AI of Alien: Isolation | AI and Games #15

Here, the alien's AI adapts to the player's situation, position, progress in the story, and proximity. Although it is not strictly speaking an AI with reinforcement learning, its operating pattern is similar and hints at some very interesting possibilities in this sector, which makes extensive use of AI programs to manage non-player characters, events, etc.

What are the prospects for AI learning?

With all that said, what conclusions can we draw and what can we expect for the future of AI learning, other than that the most exciting developments are yet to come? We can now imagine artificial intelligence with hybrid learning models, capable of combining several techniques to become more versatile and even more powerful. On the other hand, we are still far from reaching the full potential of these three learning methods, and there is still much progress to be made in AI. As it becomes more effective, more and more sectors of activity will be able to make use of it.

FAQ

1. What are the three types of learning in artificial intelligence?

Supervised, unsupervised, and reinforcement learning.

2. What is supervised learning used for?

To train an AI with labeled data for recognition, analysis, or prediction.

3. What is the difference between supervised and unsupervised learning?

The first uses data prepared by humans, while the second learns on its own to detect patterns and classify data.

4. How does reinforcement learning work?

It trains AI through a system of rewards and punishments to optimize its decisions.

5. What are the prospects for AI learning?

The future could see the emergence of hybrid models combining several methods for greater performance and versatility.