2019-12-09 20:00:00
This is a series of articles from all angles to assess a problem: "My business can not be used either to use AI AI??"
Assess the current angle - learning
Past Content:
Soul "continuous learning" is the artificial intelligence
The first two have already explained, the ability boundary little rule-based, many practical problems can not be solved by means of rules. Artificial intelligence can expand the capacity of the boundaries of the computer .
In addition to expanding the capacity of bounds, artificial intelligence and a very important feature - continuous learning, and constantly improve the capacity limit .
We all know AlphaGo victory over the world's most powerful players on the go in, but we may not know is:
AlphaGo Zero (AlphaGo upgraded version), start from blank self-Go, three days to beat AlphaGo, record is 100: 0.
In other words: the machine through sustained and rapid learning 7 × 24, just three days time will be able to accumulate beyond human than ten years.
Some extreme examples AlphaGo, learning speed of the machine will not be so much faster scene, the focus I want to express is:
In the past, the upper limit of human capabilities of the computer is given, the human need to tell the computer how to do the job.
Now, artificial intelligence can be self-taught, it is not bound by the boundaries of human cognition.
Future, fears the machine will be fully beyond human, and even pose a threat to humans.
In fact, a lot of artificial intelligence in the field of image recognition, face recognition, voice recognition beyond
For much of the problem of artificial intelligence capabilities, human beings have to a good level of artificial intelligence points:
- Currently we see are "weak artificial intelligence";
- When AI, like humans, can do a lot of things when it reached a "strong AI";
- When AI capabilities in all aspects of far more than human, to achieve a "super artificial intelligence."
So: let the machine continuous learning, artificial intelligence is the soul. Want to use artificial intelligence techniques to solve practical problems, you must consider two questions:
- I need to solve the problem is dynamic it? You need the ability to continue to learn from?
- Can I let the artificial intelligence to achieve continuous learning (below talk about the content)?
How to make the machine continuous learning?
Want to make the machine the ability to achieve continuous learning, it requires two conditions:
- Can I get feedback data?
- Whether the data can form a closed loop?
Can I get feedback data?
We think about how a child is learning to read, and a high probability start to go wrong, wrong every time when parents and teachers will tell us where wrong, what should be. It is in this "action - Feedback - correction - then act" cycle to achieve an effective learning.
With human-like literacy learning process, the machine also requires "effective feedback" to achieve continuous learning, if there is no feedback data, then the problem areas would have been problematic, can never progress.
Therefore, effective feedback data is an important part of learning, to solve the "study" the issue.
Whether the data can form a closed loop?
When we can get feedback data, machine learning can be achieved, the next step is to solve the problem of "sustainable" in.
Examples of literacy or move. If there are two children and began to learn to read.
- A child can have a teacher tutoring at any time to correct mistakes
- B children one day each week can only accept the teacher's guidance
No doubt, it must be faster A children learn better.
Machine too, so that data form a closed loop, it is hoping to get feedback data "in real time", with the children to get personal counseling as A.
So-called closed-loop data is to top "action - Feedback - correction - then act" cycle automatically on the machine operation, all without human participation.
There is a very typical example is the electronic business platform in the recommendation system in practical applications. You just read a pair of basketball shoes, the moment will give you recommend a lot of basketball shoes.
Therefore, the data form a closed loop, allowing the machine to achieve a continuous learning, to address is "sustainable" issue.
case study
Google's album Face recognition is how to collect data and feedback data to form a closed-loop?
Album App Now many have face recognition function that can help you in different photos classification of people do. However, actual use will encounter many judgment errors.
If you do not collect feedback data, so that data form a closed loop, the error will go on forever!
Step one: take the initiative to ask the user
Some photos are not high definition, or some great hairstyle change, or some people unloaded the makeup ...
There are many causes of machine doubt their own judgment, do not blame this machine, many women dressed up their loved ones would not recognize them.
This time, Google album will take the initiative to ask the user whether the two picture is the same person, as follows:
Step two: the ability to upgrade
If I chose the "same person" after the machine to collect feedback from users, the first thing is to their photos merged together.
There is one more important thing, machine learn something:
- Machine knows the make-up: So after this man make-up like this.
- Machine knows the fat: the original gain weight after things turned out.
- I know the machine age: old people will have had these changes.
Artificial intelligence tells us that Spider-Man always becomes this:
to sum up
Artificial intelligence is the reason why we have high hopes, there is a very important reason is this:
AI through continuous learning, to break through the upper limit of human capabilities. Some estimates are even machines will surpass human beings in all respects.
In order for the machine to achieve continuous learning, we need to implement two conditions:
- Constantly get feedback data, let the machine know where the good, where the bad
- The addition of the closed-loop feedback data, the sustainability of machine learning, to enhance the ability
