Machine Learning Knowledge Framework

Today began to study Stanford University CS229 course, do not want to completely copy the handout, hope to add their own understanding, beginners, inevitably error, welcome correct.

The first lesson introduces the knowledge framework for machine learning, CS229, as long as the inductive reasoning approach in machine learning is described, which induces general conclusions from enough cases and, in turn, helps people solve problems related to specific cases, especially those that have not been seen before. Inductive learning is the most important means of machine learning at present.

According to the characteristics of input data, inductive learning can be divided into four categories: supervised learning, unsupervised learning, semi-supervised learning and intensive learning.
supervised learning : "Input data" and "expected output data" are known, known as "callout" data. The task is to find the input and output correspondence (function), the "actual output data" and "Expected output data" error as a criterion to evaluate the current system performance.
Example: Provide training data on "housing information" and corresponding "house price" to predict the price of a house other than the training data
Given the "tumor volume" and the corresponding "Whether it is malignant" training data, predicting whether a tumor outside the training data is malignant

Unsupervised learning : Only input data, no expected output data, called "unlabeled data". The task is to find the distribution pattern of the input data or the correlation between the different parts.
Example: Classification of images, recognition of objects in images

semi-supervised learning : Supervised learning data need to be labeled, unsupervised learning data do not need to have labels, data volume is large, the data is a heavy burden of work, so some scholars put forward semi-supervised learning. The data used is divided into two parts: a small amount of callout data and a large amount of unlabeled data. The learning process is to obtain a preliminary model by learning the labeling data, using the model to determine the expected output of unlabeled data, transforming it into labeled data, and then learning through this part of the transformed data. (Trial-and-error learning may lead to incorrect labeling, resulting in decreased learning results)

Reinforcement Learning : According to the input data output, the rewards and punishments, rewards and punishments feedback to the model, the model to make adjustments. Reinforcement learning emphasizes cumulative benefits and focuses on the results of a series of decisions, rather than on the results of each isolated point. Reinforcement learning comes from the discipline of automatic control, the earliest mainly used in robotics.
Examples: robotics, driverless cars, etc.