Classification of Machine Learning: Over the decades, research has published a wide variety of machine learning methods, which can be classified in various ways depending on the emphasis.
Classification based on learning strategies
1. Machine learning that simulates the human brain
Symbolic learning: Simulating the macroscopic psychological level learning process of the human brain, based on the principles of cognitive psychology, using symbolic data as input, using symbolic operation as a method, and using reasoning process to search in a graph or state space, the learning goal is concepts or rules, etc. The typical methods of symbolic learning include memory learning, example learning, deductive learning, analogical learning, interpretation learning, and so on.
Neural network learning (or connection learning): Simulates the micro-physiological learning process of the human brain, is based on the principles of brain and neuroscience, uses an artificial neural network as a functional structural model, uses numerical data as input, and uses numerical operation as a method. The iterative process searches in the coefficient vector space, and the learning objective is a function. Typical connection learning includes weighted value correction learning and topology learning.
2. Machine Learning Directly Using Mathematical Methods
Mainly statistical machine learning.
Statistical machine learning is based on the preliminary understanding of the data and the analysis of the learning purpose, selecting the appropriate mathematical model, formulation hyperparameters, and input sample data, according to a certain strategy, using the appropriate learning algorithm to train the model, and finally use the trained model. model to examine and predict the data.
There are three elements of statistical machine learning:
Model: Before the model is educated, its possible parameters are many or even infinite, so the achievable models are also multiple or even infinite. The team of this model is the hypothesis space.
Strategy: The criterion for selecting the model with the best parameters from the hypothesis space. The little the error (loss function) between the prediction or classification of the model and the actual circumstances, the better the model. Then the procedure is to reduce the error.
Algorithm: The method of picking a model from the hypothesis space (equivalent to finding the best model parameters). The parameter solution of machine learning is usually transformed into an optimization problem, so the learning algorithm is usually an optimization algorithm, such as the fastest gradient descent method, Newton method, and quasi-Newton method.
Classification based on learning methods
1. Inductive learning
Symbolic induction learning: Typical symbolic induction learning includes decision tree learning and example learning.
Function induction learning (discovery learning): Typical function induction learning includes neural network learning, example learning, discovery learning, and statistical learning. The learner’s brain transforms external knowledge into an internal ability so that when encountering problems, he can handle it in different ways. This process is considered to be inductive and interpretation of the repeated role in cognitive research. The interpretation method and the previous experience reflection method have a relationship.
2. Deductive learning: Learning inductive interpretation method
3. Analogy learning: Typical analogy learning includes case (example) learning.
4. Analysis learning: Typical analysis learning includes interpretive learning and macro operation learning.
Classification based on learning style
1. Supervised learning (learning with a tutor): There is a tutor signal in the input data, and the probability function, algebraic function, or artificial neural network is used as the basis function model, and the iterative calculation method is used, and the learning result is a function.
2. Unsupervised learning (unsupervised learning): There is no mentor signal in the input data, the clustering method is used, and the learning result is a category. Typical unsupervised learning includes discovery learning, clustering, and competitive learning.
3. Reinforcement learning (reinforcement learning): A learning method guided by statistics and dynamic programming techniques with environmental feedback (reward/penalty signals) as input.
Classification based on data from
1. Structured learning: It takes structured data as input and uses numerical calculation or symbolic deduction as the method. Typically structured learning includes neural network learning, statistical learning, decision tree learning, and rule learning.
2 Unstructured learning: Taking unstructured data as input, typical unstructured learning includes analogical learning case learning, explanation learning, text mining, image mining, Web mining, etc.
Classification based on learning objectives
1. Concept learning: The goal and result of learning are concepts, or learning to acquire concepts. Typical concept learning mainly contains example learning.
2. Rule learning: The goal and result of learning are rules, or learning to obtain rules. Normally rule learning mostly includes decision tree learning.
3. Function learning: The goal and result of learning are functions, or learning to obtain functions. Typical function learning mainly includes neural network learning.
4. Category learning: The goal and result of learning object categories, or learning to obtain categories. The typical category of learning mainly includes cluster analysis.
5 Bayesian network learning: The goal and result of learning is a Bayesian network or a kind of learning to obtain a Bayesian network. It can be further divided into structural learning and majority learning.