Migration Learning & self-learning

I recently read ng's deep learning tutorial and seeSelf-taught LearningSome concepts are unfamiliar. As a part of paying off the technical debt, I spent half an afternoon searching for several terms. If I want to use them later, I will try again.

Supervised Learning has been discussed in the previous blog. Here we will mainly introduce migration learning and self-learning. Supervised Learning requires a large number of training samples. At the same time, it requires that the training samples and test samples come from the same distribution. What should I do if I cannot meet this requirement? Let's see if the following learning methods can help you.

•  Transfer Learning

A problem that sometimes bothers you is the calibration of training data. This will consume a lot of manpower and material resources. In addition, machine learning assumes that the training data and test data are distributed in the same way. However, in many cases, this same distribution assumption is not satisfied. In general, training data may expire, that is, data that is difficult to calibrate should be discarded, and a large amount of new data should be re-calibrated. The goal of migration learning is to use the knowledge learned from an environment to help learning tasks in the new environment. To put it bluntly, there is only a small amount of new labeled data, but there is a large amount of old labeled data (or even valid data of other categories ), in this case, select valid data from the old data and add it to the current training data to train the new model. The original sentence is:

Transfer learning is what happens when someone finds it much easier to learn to play chess having already learned to play checkers, or to recognize tables having already learned to recognize Chairs; or to learn Spanish having already learned Italian.

The masterpiece of migration learning isBoosting for transfer learningFor more information about it, I don't know much about it.

• Self-learning self-taught Learning

Like Semi-supervised learning, self-learning currently only has a small number of training samples on hand, but there are still a large number of unlabeled samples on hand. Here is a classic example of separating an elephant from a rhinoceros. For supervised learning, we have a large number of elephant samples and rhino samples at hand. Next we will train classifier for classification, which we all know. For migration learning, we have a large number of samples of sheep and horses on hand (marked), a small number of samples of elephants and rhinoceros, next, we need to select valid samples from the sample of the goat and the horse and add them to the marked samples of the elephant and the rhinoceros respectively, and then use the supervised learning method to train the classifier. Instead of supervised learning, there are only a small number of marked samples of elephants and rhinos on hand, and a pile of unlabeled data of elephants and rhinos (note that they are either elephants or rhinos, no other species ). Semi-supervised learning uses these samples to train classifier for classification. Self-learning is also a marked sample with only a small number of elephants and rhinos on hand, and there are a lot of natural images. The so-called natural images are images of elephants and rhinos, and images of various other species. Self-learning is more suitable for actual scenarios than semi-supervised learning. Which of the following images only have elephants and rhinos? Natural Images are more widely sourced and can be downloaded from the Internet.

Shows how to implement self-learning. First, we extract a set of features from unlabeled natural images (such as sparse dictionaries and amazing things, which will be studied later ). In this way, any labeled or unlabeled image can be expressed using this set of features. Then we train a classifier for classification (? I still don't understand. The number of training samples hasn't changed much, Todo ).

// Todo