Abstract: * * ** There are many popular words about ARTIFICIAL intelligence in our ears, such as Deep Learning, Reinforcement Learning, Transfer Learning. Many people are confused about the meaning of these high-frequency words and the relationship behind them. Today, we will clarify the relationship and difference between them.

I. Deep learning:

The success and development of deep learning benefits from the significant improvement of computing power and big data. After digitalization, a large amount of data can be generated, and the law of data can be discovered through a large amount of data training, so as to realize data prediction based on supervised learning.

Neural network-based deep learning is mainly applied to image, text, speech and other fields.

At NIPS 2016, Ng gave a technical roadmap for the future direction of AI:

Supervised learning is the AI technology with the most commercial scenarios and the highest maturity at present, and the next commercial AI technology will be Transfer learning, which is also the AI technology that Andrew predicted is most likely to be commercialized in the next five years.

Ii. Transfer Learning:

** Transfer learning: how to transfer the learned knowledge from one scene to another by training related and similar data, realizing the generalization ability of the model itself through transfer learning.

Take image recognition, from day to night, winter to summer, from identifying Chinese to identifying foreigners…

Borrow A schematic diagram (From: A Survey on Transfer Learning) to illustrate:

The value of transfer learning is reflected in:

1. The data of some scenarios cannot be collected at all, so transfer learning is very valuable;

2. Reuse of existing knowledge domain data, so that a large amount of existing work will not be completely discarded;

3. There is no need to re-collect and calibrate huge new data sets at great cost;

4. For rapidly emerging new fields, can quickly migrate and apply, reflecting the advantages of timeliness;

Summary of the practice of transfer learning algorithm:

1. Confusion training through original data and a small amount of new field data;

2. Split the original training model and retain the basic model (data) as the migration basis of the new field;

3. Get new scene images through 3D simulation (OpenAI Universe uses racing games for training);

4. The method of transfer learning with adversarial GAN;

3. Reinforcement learning:

Reinforcement Learning: Full name is Deep Reinforcement Learning (DRL), so that the machine has the ability of self-learning and self-thinking.

Reinforcement learning is currently mainly used in the field of game AI, the most famous of which is AlphaGo. Reinforcement learning is a complex proposition, and Deepmind’s David Silver understands it as an interdisciplinary discipline:

In fact, reinforcement learning is an exploratory learning method, which can be improved through continuous “trial and error”. Different from supervised learning, reinforcement learning itself has no Label, and it cannot get clear feedback after each step of Action (in this point, supervised learning can perform Label comparison for each step. Get True or False).

Reinforcement learning is described by combination of the following elements:

Object (Agent)

Which is our intelligence theme, like AlphaGo.

Environment

The scenario in which the Agent is located — for example, the board on which go is played, and its corresponding State — for example, the current chess game.

The Agent needs to get feedback from Environment awareness (whether the current situation is more favorable for me).

Action (Actions)

What actions can be taken in each State and analyze the impact of each Action.

Reward (Rewards)

A Reward or punishment received after an Action is performed. A Reward is obtained by observing the environment.

Click to follow, the first time to learn about Huawei cloud fresh technology ~