I. What is the essence of machine learning?
For a long time, human cognition has gone through three steps: raw data – professional knowledge/experience and common sense – cognition.
Faced with the same raw data (stock market data, instrument indices, social phenomena, etc.), people with different knowledge will arrive at different conclusions; similarly, people with the same knowledge will also arrive at different conclusions when faced with different situations such as no data, a small amount of data, a large amount of data, and sufficient data (equal information game, asymmetric information game).
So, which is more important, knowledge or data? For a long period in human history, those who possessed knowledge undoubtedly held the upper hand. However, the emergence of machine learning methods has completely reversed this trend. Perhaps, in the future, "knowledge" will be worthless, while "data" will be invaluable.
The essence of machine learning lies in establishing a direct mapping between raw data and cognition, breaking free from the constraints of "knowledge." From this point forward, the way human cognition works will be revolutionized. Because, from this point forward, we may no longer need that cumbersome "knowledge."
II. Machine learning can be classified as follows
(I) Classification based on learning methods
(1) Inductive learning
Symbolic inductive learning: Typical symbolic inductive learning methods include example learning and decision tree learning.
Functional inductive learning (discovery learning): Typical functions of inductive learning include neural network learning, example learning, discovery learning, and statistical learning.
(2) Deductive learning
(3) Analogical learning: Typical analogical learning includes case (example) learning.
(4) Analytical learning: Typical analytical learning includes interpretive learning and macro operation learning.
(II) Classification based on learning style
(1) Supervised learning (learning with a tutor): The input data contains tutor signals, and the model is based on probability functions, algebraic functions or artificial neural networks. Iterative calculation methods are used, and the learning result is a function.
(2) Unsupervised learning (no-supervisor learning): The input data has no supervisor signal. Clustering methods are used, and the learning result is a category. Typical unsupervised learning methods include discovery learning, clustering, and competitive learning.
(3) Reinforcement learning: A learning method that uses environmental feedback (reward/punishment signals) as input and is guided by statistical and dynamic programming techniques.
(III) Classification based on data format
(1) Structured learning: using structured data as input and numerical computation or symbolic deduction as methods. Typical structured learning methods include neural network learning, statistical learning, decision tree learning, and rule learning.
(2) Unstructured learning: using unstructured data as input, typical unstructured learning methods include analogy learning, case learning, explanation learning, text mining, image mining, and web mining.
(iv) Classification based on learning objectives
(1) Concept learning: The goal and result of learning is a concept, or learning in order to acquire a concept. Typical concept learning mainly includes example learning.
(2) Rule learning: The goal and result of learning is rules, or learning in order to obtain rules. Typical rule learning mainly includes decision tree learning.
(3) Function learning: The goal and result of learning is a function, or in other words, learning in order to obtain a function. Typical function learning mainly includes neural network learning.
(4) Category learning: The goal and result of learning is the object class, or in other words, learning in order to obtain the category. Typical category learning mainly includes cluster analysis.
(5) Bayesian network learning: The goal and result of learning is to obtain a Bayesian network, or in other words, a kind of learning aimed at obtaining a Bayesian network. It can be further divided into structure learning and majority learning.
