Machine learning

Machine learning is the part of AI that helps machines learn to understand on their own by going through large amounts of data. Unlike pure programming where you’re talking about the computer explicitly what it should do, algorithms are used that allow the computer to interpret and learn from the data it processes to then predict patterns. The more data and information the computer is exposed to, the smarter it becomes. Machine learning works similar to a child who sees a bird for the first time. The child needs to see some more examples of birds to see the difference between birds, dogs, and cats.

For machines to learn from data, various algorithms are used to make sense of data that is passed to the machine. These algorithms fall into multiple categories, where each category groups algorithms depending on their similarities in how the machine is learning. These categories include, but are not limited to, regression-, regularization-, decision tree- and clustering algorithms.

There are three common learning styles that different training algorithms can fall into:

Supervised learning means that you have a data set where you already know the correct answer and want to create/learn a model that can then be used to make predictions from data where the correct answer is not known. Computers in this case learn in the same way as people by studying many examples of how something is done.

Unsupervised learning identifies patterns in data and tries to find similarities to divide data into categories. If the data is missing labels, it will not be able to train with supervised algorithms. Instead, using unattended learning, it is up to the algorithm to find patterns in the data to categorize them.

Semi-supervised learning is a mixture of supervised and unsupervised learning. It uses labeled and unlabeled data to learn and understand the complete dataset. Since labeled data is more expensive than unlabeled data, a combination of slightly labeled data and a large amount of unlabeled data is used instead.

An AI software called AlphaGo was developed by Google that, back in 2016, defeated a Grandmaster in the game Go, which wasn’t expected until 2026. In Go, there are up to 200 moves each turn, compared to chess where there are 20. This makes it extremely complex to calculate the most efficient move on each turn, which is what makes this feat so significant. Instead of calculating all possible moves, the software was fed data from 30 million games of Go by human players, and by feeding it into a deep learning neural network. In addition, the AI was taught by playing against itself, making it even better at the game. By playing against the AI, the Go players have said that they are now re-evaluating what they know about the game and testing out new moves that the AI used against them when playing.

Reinforcement learning

Reinforcement learning is rewarding the AI for performing a correct operation or choice while punishing the wrong choice or operation. For example, in a PacMan game: “from pixel data an agent might be given a numeric reward for the result of a unit of travel: 0 for empty space, 1 for pellets, 2 for fruit, 3 for a power pellet, 4 for a ghost post-power pellet, 5 for collecting all pellets and completing a level but being deducted 5 points for collision with a ghost.”

Regression

Regression is used to find a correlation between values in, for example, housing prices. The algorithm tries to find the most suitable formula for representing the house price — for example, using a polynomial function, price = c1 + area * c2 + area² * c3, where the algorithm tries to figure out c1, c2, and c3 for some given house prices. By testing the mathematical model best suits the data that is given, the algorithm learns how to predict a new house’s price, depending on the calculated values for c1, c2, and c3 and the area of the new house. This algorithm is supervised, since the prices have to be made sure to be accurate, it cannot be assumed that the algorithm will choose the right formula, which may end up giving an inaccurate prediction

1. https://www.mathworks.com/discovery/machine-learning.html
2. https://machinelearningmastery.com/a-tour-of-machine-learning-algorithms/
3. https://www.mathworks.com/discovery/supervised-learning.html
4. https://www.mathworks.com/discovery/unsupervised-learning.html?s_tid=srchtitle
5. https://www.datascience.com/blog/what-is-semi-supervised-learning
6. https://machinelearningmastery.com/a-tour-of-machine-learning-algorithms/
7. https://searchenterpriseai.techtarget.com/definition/reinforcement-learning
8. https://www.zdnet.com/article/googles-alphago-retires-after-beating-chinese-go-champion/