First, the definition of machine learning
Machine learning is more than one field of interdisciplinary research algorithms from the data, or computer simulations to study how to achieve human learning behavior, the algorithm selected based on existing data or past experience, build a model to predict the new data, and re-organization existing knowledge structures so as to continuously improve their performance.
Second, the classification of machine learning
1, supervised learning SL (Supervised Learning)
With some known properties of the sample (x, y) as a training set to build a mathematical model to solve f: x-> y, predict unknown samples.
1) discriminant model (Discriminative Model)
Conditional probability p (y | x) modeling, such as linear regression, decision trees, SVM, KNN, neural networks.
2) the model formula (Generative Model)
Joint probability p (x, y) is modeled as HMM, Naive Bayes, GMM, LDA and the like.
2, unsupervised learning UL (Unsupervised Leaning)
Training data only x labels is not y, the purpose of trying to extract the data and rules implicit in the structure, the common algorithm K-means, dimension reduction, text processing (feature extraction) and the like.
As a general unsupervised learning supervised learning pre-data processing function is extracted from the original data set required label information.
3, semi-supervised learning SSL (Semi-Supervised Learning)
A combination of both before, marked with a small number of samples and unlabeled samples for training and testing. Three kinds of SSL basic assumptions:
1) Suppose smoothing (Smoothness Assumption): two similar sample located in close proximity of the dense region data class label.
2) Suppose the cluster (Cluster Assumption): When the two samples are in the same cluster clusters, they have the same class of labels at high probability.
3) Suppose manifold (Manifold Assumption): embedding high-dimensional data into low-dimensional manifold, when two samples located in a small local neighborhood of low-dimensional manifolds, which have a similar class label.
4, reinforcement learning RL (Reinforcement Learning)
System ( Autonomous Agent) by learning from the environment States to conduct mapping Actions to strengthen the signal Rewards function value is maximized, that is, to improve the program of action Policy to adapt to the environment.
RL and SL, UL is an important difference: the latter two need to meet the data is subject to independent and identically distributed, and RL this condition is not satisfied, because the Agent and environmental interaction in the state is constantly changing, RL is essentially continuous the process of making decisions sequences.
Class Notes:
Derivative:
surroundings:
