The term "machine learning and data analysis" was coined in 1959 by Arthur Samuel. Machine learning explores the study and construction of algorithms that make it possible to learn and make predictions from data - such algorithms are superior to strictly defined static program commands by issuing predictions or solutions. Machine learning is used in a number of computational tasks, where the design and programming of explicit algorithms with good performance is difficult or impossible. Examples of applications include email filtering, network intruder detection, and computer vision.
The bottom line is
Machine learning is closely related (and often coincides) with computational statistics, which also focuses on forecasting using computers. It has strong links with mathematical optimization, which provides methods, theories and applications in this scientific field. Machine learning is sometimes combined with data mining, where the last subfield focuses more on data analytics and is known as unsupervised learning.
Machine learning and data analysis is a method used to develop complex models and algorithms that can be predicted. In commercial use, this is called predictive analytics. These analytic models allow researchers, scientists, engineers, and analysts to create reliable, repeatable solutions and results, and to reveal hidden ideas by exploring historical relationships and trends in data.
Algorithm Example
Tom M. Mitchell presented a widely cited, more formal definition of algorithms studied in the field of machine learning: “It is said that a computer program learns from experience E with respect to a certain class of tasks T and a performance indicator P, if its performance when performing tasks in T, measured P improves with experience E ". Letters are formal designations of algorithms. This is a definition of the tasks that machine learning and data analysis considers.
History
The emergence of this miracle science followed after the proposal of Alan Turing in his article “Computer Engineering and Intelligence”, in which the question “Can machines think?” replaced by the question: “Can machines do what we (as thinking entities) can do?” Turing's proposal reveals various characteristics that a thinking machine may possess, and the various consequences of its construction.
Arthur Samuel, an American pioneer in the field of computer games and artificial intelligence, coined the term "machine learning" in 1959. As a scientific discipline, machine learning has grown out of a desire for artificial intelligence. Already in the early days of AI, as an academic discipline, some researchers were interested in machines learning from existing data. They tried to approach the problem using various symbolic methods, as well as what was then called "neural networks." These were mainly perceptrons and other models, which were later recognized anew in generalized linear statistical models. Probabilistic reasoning and constructing models of future events were also used due to the probability of their occurrence, especially in automated medical diagnostics.
Artificial intelligence problem
However, a growing emphasis in a logical, knowledge-based approach has caused a gap between AI and machine learning. Probabilistic systems suffer from theoretical and practical problems of data collection and presentation. By 1980, expert systems began to dominate AI, and statistics were not in favor of artificial intelligence, which was and remains too imperfect. Work on symbolic / knowledge-based learning continued within the framework of AI, which led to inductive logical programming, but a more statistical line of research is currently beyond the scope of AI itself in pattern recognition and information retrieval. Research on neural networks was abandoned by AI and computer science at about the same time. This line also continued outside the AI / CS field as a link between researchers from various disciplines, including Hopfield, Rumelhart and Hinton. Their main success was achieved in the mid-1980s with a rethinking of backpropagation.
The specialization Machine Learning and Data Analysis, reorganized as a separate discipline, began to flourish in the 1990s. At the moment, the goal of this discipline is to achieve the creation of artificial intelligence to solve solvable problems of a practical nature. She shifted the focus from symbolic approaches that were inherited from the first experiments with AI to methods and models borrowed from statistics and probability theory.
Data mining
Judging by the reviews, machine learning and data analysis courses often use ready-made data, with which you can develop computers and mechanisms, making them a distant likeness of artificial intelligence. Data mining, in turn, focuses on detecting previously unknown properties in the data (this is the stage of analyzing the discovery of knowledge in databases). For data mining, many machine learning methods are used, but with different goals. Machine learning, on the other hand, also uses data mining techniques as “unsupervised learning” or as a preprocessing step to improve learning accuracy.
Most of the confusion between the two research communities (which often have separate conferences and separate journals, ECML, PKDD, which is the main exception) comes from the basic assumptions that they work with: in computer training, effectiveness is usually assessed in terms of the ability to reproduce known knowledge, and When opening knowledge and data mining (KDD), the key is to detect previously unknown knowledge. Evaluated with respect to known knowledge, an uninformed (uncontrolled) method will easily surpass other controlled methods, while in a typical KDD problem controlled methods cannot be used due to the inaccessibility of training data.
Optimization
Machine learning also has a close connection with optimization: many learning problems are formulated as minimizing losses during the learning process itself. The loss functions express a mismatch between the predictions of the model being trained and the actual results.
Machine Learning and Data Analysis: MIPT
The main goal of the student is to summarize his experience. This also applies to training at MIPT, in which a machine learning course is available. There, students are trained using educational computers to accurately perform new, unprecedented examples / tasks after they analyze the starting data set. The examples solved during the training are taken from some well-known probability distribution, and the student must build a general model about this space, which allows him to make fairly accurate predictions in new cases.
Algorithm Analysis
Computational analysis of machine learning algorithms and their efficiency is a branch of theoretical informatics known as the theory of computational learning. Since learning sets are finite and the future uncertain, learning theory usually does not guarantee the execution of algorithms. Instead, probable performance estimates are fairly common. Displacement-dispersion decomposition is one way to quantify the generalization error.
Data complexity
To achieve maximum performance in the context of generalization, the complexity of the hypothesis must correspond to the complexity of the function that underlies the data. If the hypothesis is less complex than the function, then the model does not match the data. If the complexity of the model increases in response, then the learning error decreases. But if the hypothesis is too complicated, then the model is subject to redefinition, and generalization will be worse. And this conclusion we draw from many journals of machine learning and data analysis that persist after serious scientific work and research in this area.
In addition to performance limitations, computational theorists study the time complexity and feasibility of learning. In the theory of computational learning, computation is considered feasible if it can be done in polynomial time. There are two kinds of time complexity results. Positive results show that a certain class of functions can be studied in polynomial time. Negative results show that some classes cannot be studied in polynomial time. Therefore, for those who do not know where to start machine learning and data analysis, there are special data analysis tasks that are available in courses at many universities.