Source: Deep Learning on Medium

Machine Learning (ML) is a scientific study of algorithms and mathematical models used by computer systems to perform a specific job without explication. It is regarded as an artificial intelligence branch. Machine learning algorithms create a sample information-based mathematical model, known as “training data,” to make predictions or decisions without the function being specifically programmed. Machine learning algorithms are used in a wide range of applications, such as email sorting and computer vision, where designing a traditional algorithm is impossible or unfeasible to perform the task effectively.

Machine learning is closely related to mathematical statistics which concentrate on computer-based predictions. The analysis of mathematical optimization provides the area of machine learning techniques, theory and application areas. Data mining is an area of study that fosters exploratory data analysis through unregulated learning.

There are several broad categories of machine learning tasks. The algorithm constructs a mathematical model in supervised learning from a set of data containing both the inputs and the desired outputs. For example, if the task were to decide whether an image contained a certain object, images with and without that object (the input) would be included in the training data for a supervised learning algorithm and each image would have a tag (the output) indicating whether it contained the object. The data may only be partly accessible and restricted to direct feedback in special cases. Semi-supervised learning algorithms create mathematical models from incomplete training data, where there are no labels for a portion of the sample information.

Supervised learning types are classification algorithms and regression algorithms. Classification algorithms are used to restrict the outputs to a specific set of values. The input would be an incoming email for a classification algorithm sorting messages, and the output would be the name of the directory to file the email in. For an algorithm detecting spam emails, the result would be either “spam” or “not spam” prediction, represented by true and false Boolean values. Regression algorithms are named for their continuous outputs, which means they can have any value within a range. Examples of a continuous value are the temperature, duration or price of an object e.t.c.

In unsupervised learning, the algorithm builds a mathematical model from a data set containing only inputs and no desired output labels. Unsupervised learning algorithms are used to locate data structure, such as data point clustering or clustering. Unsupervised learning can discover patterns in data and can group inputs into categories, such as feature learning. Reduction of dimensionality is the process of reducing the number of “features” or inputs in a data set.

Active learning algorithms control the desired outputs (training labels) for a limited set of budget-based inputs and optimize input selection for which training labels will be acquired. These may be introduced to a human consumer for marking when used interactively. In a dynamic environment, reinforcement learning algorithms are provided with feedback in the form of positive or negative reinforcement and are used in autonomous vehicles or in learning to play a game against a human opponent. Many advanced machine learning algorithms include topic modeling, where a collection of natural language documents are given to the computer program and other documents covering similar topics are found. Machine learning algorithms can be used in density estimation problems to find the unobservable probability density function. Based on the previous experience, meta-learning algorithms learn their own inductive bias.

# Relation to data mining

Machine learning and data mining often use the same methods and greatly overlap, but while machine learning focuses on prediction based on known properties learned from training data, data mining focuses on discovering (previously) unknown properties in the data (this is the step of evaluating the discovery of information in databases). Data mining uses many methods of machine learning, but with different objectives; on the other hand, machine learning often uses methods of data mining as “unsupervised learning” or as a pre-processing step to improve learner accuracy. Much of the ambiguity between these two research groups (which often have different conferences and separate publications, ECML PKDD being a significant exception) arises from the basic assumptions with which they work: in machine learning, quality is generally measured with regard to the ability to reproduce existing information, while in knowledge discovery and data mining (KDD) the main task is to discover knowledge.

# Relation to optimization

Machine learning also has close links to optimization: most learning problems are constructed on a training set of examples as a minimization of some loss function. Loss functions describe the difference between the expectations of the model being trained and the actual problem instances (for example, one needs to assign instances to a label in classification, and models are trained to correctly predict a set of examples’ pre-assigned labels). The difference between the two fields stems from the objective of generalization: while optimization algorithms can minimize the loss on a training set, the purpose of machine learning is to minimize the loss on unseen samples.

# Relation to statistics

Machine learning and statistics are methodologically closely related areas, but distinct in their main objective: statistics draw population inferences from a sample, while machine learning seeks predictive patterns that are generalizable. The theories of machine learning, ranging from scientific concepts to analytical methods, had long pre-history in statistics, according to Michael I. Jordan. He also proposed naming the overall sector for the term data science as a placeholder. Leo Breiman differentiated two paradigms of statistical modeling: data model and algorithmic model, in which “algorithmic model” means more or less machine learning algorithms such as Random Forest. Machine learning approaches have been implemented by some statisticians, contributing to a new area they call statistical learning.

To learn more about machine learning or its application in your field, follow the link by clicking here