Abstract: Many popular vocabularies about artificial intelligence linger in our ears, such as Deep Learning, Reinforcement Learning, and Transfer Learning. Many people are concerned about the meaning of these high-frequency vocabulary and their backing I’m confused about the relationship, and today I will clarify the relationship and difference between them.
1. Deep learning:
The success and development of deep learning benefited from the significant increase in computing power and big data. After digitization, a large amount of data is generated. A large amount of data training can be used to discover the laws of the data, so as to realize data prediction based on supervised learning.
Deep learning based on neural networks is mainly used in image, text, voice and other fields.
At the 2016 NIPS conference, Wu Enda gave a technical development map for the future direction of AI:
Supervised learning is currently the most commercially available AI technology with the highest maturity. The next commercial AI technology will be Transfer Learning. This is also the AI that Andrew predicts is the most likely to be commercialized in the next five years. technology.
2. Transfer learning:
Transfer learning: Use relevant and similar data to train and realize the generalization ability of the model itself through transfer learning, which is how to transfer the learned knowledge from one scene to another.
Take image recognition, from day to night, from 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. Data in some scenes cannot be collected at all, at this time transfer learning is very valuable;
2. Reuse existing knowledge domain data, and a large amount of existing work will not be completely discarded;
3. There is no need to spend a huge price to reacquire and calibrate a huge new data set;
4. For rapidly emerging new areas, it can be quickly migrated and applied, reflecting the advantages of timeliness;
Summary of the practice of transfer learning algorithms:
1. Obfuscate 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 for the new field;
3. Get new scene images through 3D simulation (OpenAI's Universe platform uses racing games to train);
4. The method of migration learning with the help of confrontation network GAN;
3. Reinforcement learning:
Reinforcement learning: The full name is Deep Reinforcement Learning (DRL), which gives machines the ability to learn and think on their own.
At present, reinforcement learning is mainly used in the field of game AI, and the most famous one should be AlphaGo's Go War. Reinforcement learning is a complicated proposition. Deepmind god David Silver understands it as such an interdisciplinary subject:
In fact, reinforcement learning is an exploratory learning method that is improved through continuous "trial and error". The difference from supervised learning is that reinforcement learning itself does not have a label, and it cannot get clear feedback after each step of Action (in At this point, each step of supervised learning can perform Label comparison and get True or False).
Reinforcement learning is described by combining the following elements:
Object (Agent)
That is our smart theme, such as AlphaGo.
Environment (Environment)
The scene where the Agent is located-such as the board where Go is played, and its corresponding state-such as the current game.
Agent needs to get feedback from Environment perception (whether the current situation is better for me).
Actions
What actions can be taken in each State, and analyze its impact for each Action.
Rewards
The reward or punishment obtained after the action is executed, the reward is obtained through observation of the environment.
Click to follow and learn about Huawei Cloud's fresh technology for the first time~