Machine Learning

 Thanks to machine learning, computers are now able to predict weather, determine stock market results, understand shopping habits, control robots in a factory, etc. they are “trainable” for these situations. Companies such as Google, Amazon, Facebook, Netflix, LinkedIn support all of the more popular consumer-oriented services with machine learning. But at the heart of all this learning is what is known as an algorithm. In summary, an algorithm is not a complete computer program, but a finite sequence of steps to solve a single problem. Certain steps are taken in the algorithm to achieve a specific and single goal.

In essence, machine learning is based on trial and error. A program that can help a self-driving car distinguish a pedestrian from a tree or a vehicle may not be written manually, but an algorithm can be created for a program that can solve this problem using the data. Algorithms can also be created to help programs predict the path of a hurricane, diagnose early Alzheimer’s, identify the world’s most paid and unpaid football stars. Machine learning algorithms basically enable programs to make predictions and get better results in these predictions based on trial and error experience over time.

There are four main types of machine learning algorithms:

• Supervised Learning,
• Unsupervised Learning,
• Semi-supervised Learning,
• Reinforcement Learning.

A computer program containing data labeled in supervised learning is provided. For example, when the task of separating cat and dog images is defined using an algorithm for sorting images, those with cats will have a “cat” tag and images with dogs will have a “dog” tag. This is considered a “training” dataset and the tags remain in place until the program can successfully sort the images at an acceptable rate. At the end of the process, the function that best describes the input data is selected and makes the best estimate “y” (output) for the given “X” (input). Supervised learning algorithms try to model the relationships and dependencies between target prediction output and input properties in such a way that they can predict output values for new data based on the relationships they have learned from previous data sets.

• Nearest Neighbor,
• Naive Bayes,
• Decision Trees,
• Linear Regression,
• Support Vector Machines,
• Artificial Neural Networks.

algorithms such as are the main types of supervised learning algorithms.

Unsupervised learning does not contain any tags. Instead, the program blindly throws in the task of splitting the cat and dog images into two groups using one of two methods. In the clustering algorithm, leg length, body length, eyes, etc. Similar objects are brought together based on properties such as. The other algorithm is called association and rules are created based on the similarities the program discovers. In other words, a common pattern is determined among the images and the images are sorted accordingly. It is a family of machine learning algorithms mainly used in pattern detection and descriptive modeling.

• K-Mean Clustering,
• Association Rules.

algorithms such as unsupervised learning algorithms are the main ones.

Semi-supervised learning is a learning paradigm related to studies on how natural systems like humans learn in the presence of both tagged and untagged data [1]. Only a few pictures are tagged in semi-supervised learning. The computer program then uses an algorithm to make the best guess about the untagged images, and then the data is fed back into the program as training data. A new set of images with just a few tags is then presented. The program is a repetitive process until it can distinguish between dogs and cats at an acceptable rate. Semi-supervised learning is in between the previous two. Labeling is very costly as it requires experts in many cases. Thus, semi-checked algorithms are the best candidates for model building, although most observations are present in the absence of tags but in small numbers. These methods take advantage of the notion that although group membership of unlabeled data is unknown, these data carry important information about group parameters.

In reinforced learning, the program knows the rules of the game and how to play it, and implements the steps to complete the round. Chess could be an example of such an algorithm. The only information given to the program is whether or not he won the match. He keeps repeating the game, keeping track of his successful moves until he finally wins a match. It constantly learns from the environment iteratively. In the process, the tool learns from its experiences with the environment until it investigates all possible situations.

• Temporal Difference (TD),
• Q Learning,
• Sarsa algorithms.

are the main reinforced learning algorithms.

I will refer to the content of these learning types in depth in the following posts.

REFERENCES

[1] Zhu, X., and Goldberg, A. B. (2009). Introduction to Semi-Supervised Learning. Artificial Intelligence and Machine Learning, 130. doi: 10.2200/S00196ED1V01Y200906AIM006

[2] Savaş, S. (2019), Karotis Arter Intima Media Kalınlığının Derin Öğrenme ile Sınıflandırılması, Gazi Üniversitesi Fen Bilimleri Enstitüsü Bilgisayar Mühendisliği Ana Bilim Dalı, Doktora Tezi, Ankara.