Google scientist Jeff Dean: How far is artificial intelligence away from popularity?
Q: What are the main challenges faced by researchers in the process of promoting research in the field of artificial intelligence?
Human learning has a lot of content from unsupervised learning, that is, you are only observing the world around you and understanding the truth of things. This is a very active field of machine learning research, but the current research progress and supervised learning are still incomparable.
In other words, unsupervised learning refers to a person's learning through observation and perception. If the computer can observe and perceive itself, it can help us solve more complicated problems.
Yes, human insights are mainly trained through unsupervised learning. You will observe the world from an early age, but occasionally you will get some signals of supervised learning. For example, someone will tell you: "That is a giraffe" or "That is a car." Once you have obtained a small amount of supervised information, your mental model will naturally respond to it.
We need to more closely integrate supervised and unsupervised learning. However, judging from the working mode of most of our machine learning systems, we have not fully advanced to that point.
Can you explain what is "enhanced learning" technology?
The idea behind "Reinforcement Learning" is that you don't necessarily understand the actions you might take, so you'll try a series of actions you should take first. If you think an idea is good, you can try it first, then Observe the reaction from the outside world. This is like playing a board game, you can respond to your opponent's actions. Eventually after a series of similar behaviors, you will get some kind of reward signal.
The idea of ​​reinforcement learning is that you can assign credit or fault to all the actions you take during the trial process while you receive the reward signal. This technology is indeed very effective in some areas of today.
I think some of the challenges of intensive learning are focused on when the behavior you can take is extremely broad. In the real world, humans can take a very wide range of behaviors at any given time. And when you play a board game, you can only take a limited set of behaviors, because the rules of the game limit you, and the reward signal is much clearer – not winning or losing.
If my goal is to make a cup of coffee, then the potential behavior I might take is quite broad, and the reward signal is not so obvious.
But you can still break the steps, right? For example, if you want to make a cup of coffee, you can learn that if you don't fully grind the coffee beans before brewing, the coffee you brew will not taste good.
Correct. I think one of the characteristics of enhanced learning is that it needs to be explored, so it is often difficult to use it in a physical system environment. But we have already begun to try to use this technology on the robot. When a robot needs to take certain actions, the behavior it can take on a particular day is limited. But if you use computer simulation, you can easily use a large number of computers to get millions of samples.
Has Google begun to use reinforcement learning technology in its core search products?
We have applied reinforced learning technology to our core products through our joint efforts with DeepMind, a startup in artificial intelligence (acquired by Google in 2014) and our data center operations staff. They also applied this technology to the air conditioning temperature control system in the data center, which achieved the same, safe cooling effect and operating conditions while greatly reducing energy consumption. It can explore which setting of the temperature control knob is reasonable and how it should respond when you change the operating conditions.
Through intensive learning techniques, they are able to explore the optimal settings for these 18 or more temperature-controlled knobs, which may not have been done by staff dedicated to temperature control. People who are familiar with the temperature control system may think: "This setting is really strange." However, in fact, it works very well.
What kind of task is more suitable for applying reinforcement learning technology?
The case of the data center mentioned above works well because there is not much different behavior in a given period of time. The temperature control system has about 18 temperature control knobs, you can turn a knob up or down, and the results are easy to measure. As long as you operate within an acceptable temperature range, your energy efficiency will be better. From this perspective, this is almost an ideal use case for reinforcement learning techniques.
korlen electric as a Isolator switch manufacturer in china, specialized in manufacturing Isolator switch in wholesale with high quality.
Do you need some Isolator switch? Our company, GAONENGGELE ELECTRICAL SHARES CO.,LTD. is the best manufacturers\factoy\suppliers in china who provided Isolator switch can help you.
China Isolator switch,Isolator switch,DIN Rail Isolator,Isolator
Wenzhou Korlen Electric Appliances Co., Ltd. , https://www.zjmoldedcasecircuitbreaker.com