Machine Learning: Learning Paradigms
Hello, in this article I am going to talk about the learning paradigms that form the basis of Machine Learning.
Machine learning algorithms are the method that reveals the pattern on the data points.
Various predictions, classifications and segmentations can be made with the pattern found. When creating a machine learning algorithm, we search for a suitable model based on the state of the data. But first of all, you should know that machine learning models are divided into various classes according to their learning methods. This distinction is differentiated according to the output of the model and according to the learning styles.
Note: Only basic learning paradigms are included in this article, learning paradigms are not limited to those in this article.
Let’s examine learning paradigms.
Supervised Learning
Most machine learning algorithms use supervised methods. Its main purpose is to establish a semantic relationship between the input variables and the output variable through a mathematical model. In these methods, we usually divide the data into train and test before giving it to the algorithm. The Train set allows modeling the mathematical relationship between input and output variables. Then, how well the created model works with the test set is tested. In summary, the labels corresponding to the new features are predicted thanks to the pattern between the features we have and the tags corresponding to these features. Classification and Regression methods are examples of this learning method. As the regression output, numerical real data emerges and this value is tried to be estimated directly. Classification methods expect a class as output, such as ‘Female’, ‘Male’, ‘Tall’, ‘Short’.
Semi-Supervised Learning
Bringing together the data used in a machine learning algorithm is the result of a long process. Labeling aggregated data is a difficult and costly journey. This problem has led to the idea of creating a model by using labeled and unlabeled data together. And as a result, the semi-supervised learning method has emerged.
In this learning method, by using labeled and unlabeled data together, unlabeled ones are estimated and train data is created.
Unsupervised Learning
In this type of learning, all data is ineffective. Data is correlated via distance and density metrics. The data is divided into clusters according to the relationship between them. It reduces the need for human intervention. Clustering and size reduction methods Clustering and dimension reduction methods are examples of unsupervised learning algorithms.
Reinforced Learning
Reinforced Learning is a trendy topic nowadays.
“Can machines learn by themselves?”
Reinforced Learning emerged with this question. The question does not mean that the machine brutally learns everything that comes it’s away. Reinforced Learning models are where the machine develops what we have taught it, rather than learning new things. The main purpose is to improve the machine itself towards a defined target. Agent and Reward concepts form the working principle. Let’s explain this;
For example, if you want to teach your dog behavior, you first imitate the behavior and show it to your dog, and then you expect your dog to do the same. If the dog exhibits the behavior you want, you reward it with food. We can expect him to do it with more enthusiasm and advancement next time. Reinforced Learning models also work with this logic. A reward for positive action and the principle of displaying the more advanced of that action in the next move.
For example, we can examine AlphaGO, which is frequently given under this title. AlphaGo is an artificial intelligence-based GO player developed by the Deepmind team. When it started to be developed, a number of GO competitions were taught to this structure, and it produced pretty bad things at first. But over time, he accumulated inferences such as “then I should do this” and demonstrated his success by beating European GO champion Fan Hui in 2015.
