1: Machine learning classification 1 (according to different uses)
1. Supervised learning:
(1) Classification: label is discrete
(2) Regression: label (the predicted value is more appropriate here, such as house price, the angle of steering wheel rotation in driverless driving) is continuous
In some cases, continuous predictions can be quantified, thereby transforming a regression problem into a classification problem.
2. Unsupervised learning:
(1) Dimensionality reduction of the data:
Feature extraction:
Feature compression: PCA, which turns a feature with a high degree of association (similarity) into one.
For example, in the recognition of handwritten numbers, the gray values of adjacent pixels are similar, so one pixel can be used to replace the surrounding ones, thereby reducing the number of features and reducing the dimension.
The meaning of data dimensionality reduction: data visualization. For example, the high dimension is reduced to two-dimensional and three-dimensional, which is convenient for drawing, and the specific distribution of data is displayed, so that people can understand it better.
(2) Anomaly detection
3. Semi-supervised learning
In practice, the label value of some samples (which can also represent the predicted value in regression) may be lost, some samples have labels, and some have no labels, which is called semi-supervised learning.
Semi-supervised learning method: Generally, unsupervised learning methods are used to process data first, and then supervised learning methods are used for model training and prediction.
Two, machine learning classification two (according to whether it can adapt to changes in the environment)
1. Batch learning:
(1) Definition: Batch learning means that after the model is trained with a batch of data, the training data is discarded. Input samples in the real environment, predict samples, and the model is no longer updated.
(2) How to adapt to changes in the environment?
Re-batch learning is performed regularly, but the amount of computation is large, and some businesses may require frequent re-batch learning.
2. Online learning;
(1) Definition:
It can reflect new data changes in a timely manner, and is also suitable for environments with large amounts of data that cannot be fully learned in batches.
(2) What if the new (abnormal) data brings bad influence?
Strengthen the abnormal monitoring of data (use the abnormal monitoring function in unsupervised learning)
three. Machine learning classification three (divided according to parameters)
1. Parameter learning
Assume first: f(x)=ax+b, according to the coordinates (samples) of the points, training, learning to get the values of the parameters a, b
2, Non-parametric learning ( I don't understand this very much, and it will be updated later )
Don't make too many assumptions about it, but that doesn't mean it doesn't have parameters.