几种不常见的machine learning: one class/reinforcement/few shot/example based learning

one class learning

Why important: hacker problem
Why is SVDD is a famous one-class algorithm: Generally, we only have one-class data, so SVDD is a good way to make a boundary to include these data.

reinforcement learning

Should be solved: creative work
Could be solved: have no certain label (correct answer) chat bot can be solved by supervised learning and reinforcement learning.
Not suitable to be solved: true or false problem. certain label. for example, face recognition.
Relationship: between classification/regression/ranking/reinforcement learning
Assume that we have the result type 1/2/3, then

• for classification, 1/2/3 means the classification type, and there is no difference between them.
• for regression, 1/2/3 is the numerical result, which means the difference between 1 and 2 equals that between 2 and 3. Moreover, the distance from 1 to 3 is larger than that from 1 to 2.
• for ranking, what we know is the order of 1/2/3, however, how much is the distance of 1 to 2, as well as the distance of 2 to 3 are unknown. we also do not know if they are same or not.
• For reinforcement learning. 1/2/3 are the indicators of the goodness or how much the reward is of one operation, such as turn left or turn right. It is used to decide what the system do at the next step.

few shot learning

How to define few: It depends. How many categories does the data have? How many features does the data have?
Spare learning and few shot learning: Spare learning is a technical method in the situation of limited computing resource. However, few shot learning is a way to using few samples to learn the knowledge without computionality limitation. Actually, few-shot learning is much likely to human beings (image that how a baby recognize dogs and cats?), thus some people claim that it is the way to general artificial intelligence.

example based learning

Rule-based method: Automatic theorem proving
Example-based method: neural network
Rule-based method to example-based method: Application-driven. even though almost all example-based learning is not explainable (such as deep learning), but they can achieve a better performance. Just like a software, we need not to know how it works on the computer, and what we need to know is know to use it.
example-based method to Rule-based: Due to mass media and business publicity stunt，many people and government are imaging a smart furture, such as auto driving, smart home, intelligent city… etc. However, better performance does not mean perfect performance. In some cases, any defect will cause huge damage, for example, auto driving. In deep learning, when you encounters a accidental problem and want to find the reason, but you may face hundreds and thousands nodes or layers. Here, people want to call rule-based learning back. When performance up to bottle neck, rule-based learning will be appear and figure out why. But it is noted that, not all example-based learning is not explainable, such as decision tree.