Machine Learning: Introduction and Approaches
We all know that computers are designed to perform tasks that are given to them in the form of commands or instructions in a program. But would you believe me if I told you computers can now learn by themselves to predict or calculate results after training themselves with test data sets? Unbelievable? But it is true. This learning done by computers without being given explicit commands is aptly called Machine Learning. The need for machine learning arises in complex situations, where it is not possible to design algorithms manually to solve problems. The computer trains itself to learn from data available and develop its own models/rules to efficiently and effectively accomplish tasks. Machine Learning is an application of Artificial Intelligence, built on fundamentals of computer science, mathematics and statistics. It is the study of algorithms that can build mathematical models on test data and use them on raw data to get results or make predictions. The learning process involves working with test data which shows examples or correct choices for given inputs, looking for patterns, general outlines and making effective predictions based on them.
Machine Learning approaches:
Supervised learning: Supervised algorithms apply the learning from past to new data. A training dataset is fed so the algorithm can create mathematical models to predict the correct or expected outputs. The learning algorithm is hence ‘trained’ and is able to correct any errors in its predictions by comparing them with the expected results. Later the trained algorithm is used with new sets of data. Example for use of this algorithm is weather forecasting. Here a lot of information like historical temperature, humidity, precipitation data is fed in. Using this information, a forecast of temperature and weather conditions for a given day can be made.
Unsupervised learning: These algorithms are used if the data available is unlabeled or unstructured. These algorithms create mathematical models to detect hidden patterns or structures in data. There is no training data and correct/expected output given to ‘train’ the algorithms. Unsupervised learning is beneficial in exploratory analysis and anomaly detection. Example for use of this algorithm is the recommender system. Amazon recommender system suggests products to purchase based on browsing and purchase history.
Reinforcement learning: These algorithms are given rewards instead of labels to assist learning. The way these algorithms learn is the closest to the way humans learn. The rewards work on the logic of reinforcing positive behaviors and punishing negative behaviors. The algorithm interacts with dynamic data rather than static data to achieve a goal. As the algorithm learns, it is provided with appropriate feedback to ensure the goal can be achieved. Example for use of this algorithm is self-driving cars. The initial position and actions that can be followed by the algorithm are fed in. The algorithm that drives these cars needs to maneuver in a constantly changing environment and receives rewards based on favorable decisions.
Picture Source: towardsdatascience.com
