This is a series of articles that evaluate a question from various perspectives: "Do I need AI for my business? Can I use AI?"
Evaluation perspective in this issue-learning
List of article series:
`` Continuous learning '' is the soul of artificial intelligence
The previous two articles have explained that rule-based capability boundaries are small, and many practical problems cannot be solved by rule methods.Artificial intelligence can expand the boundaries of computer capabilities.
In addition to expanding the capability boundary, artificial intelligence also has a very important characteristic-Continuous learning and continuous improvement.
Everyone knows that AlphaGo defeated the world's most powerful Go masters on Go, but what you may not know is:
AlphaGo Zero (an upgraded version of AlphaGo), self-taught Go from a blank, 3 days defeated AlphaGo, the record is 100: 0.
That is to say: the machine can surpass the accumulation of more than ten years of human beings in only 7 days through continuous and rapid learning of 24×3.
The AlphaGo example is a bit extreme. In many scenarios, the machine learning speed will not be so fast. The points I want to express are:
In the past, the upper limit of computer power was given by human beings, and human beings were required to tell the computer what to do.
Now, artificial intelligence can be self-taught and not bound by the boundaries of human cognition.
In the future, people worry that machines will overtake humans in all aspects and even threaten humans.
In fact, artificial intelligence has surpassed many areas such as image recognition, face recognition, and speech recognition.
Regarding the question of how big the artificial intelligence ability is, human beings have given a good grade to artificial intelligence:
- At present, everyone sees "weak artificial intelligence";
- When AI, like humans, can do many things, it reaches "strong artificial intelligence";
- When AI's capabilities in all aspects have far surpassed humans, "super artificial intelligence" has been achieved
So: letting machines continue to learn is the soul of artificial intelligence. To use artificial intelligence technology to solve practical problems, you must consider 2 problems:
- Is the problem I need to solve dynamically changing? Need the ability to continue learning?
- Can I make artificial intelligence for continuous learning (see below)?
How to make the machine continue to learn?
In order for a machine to achieve continuous learning, it needs to have 2 conditions:
- Can I get feedback data?
- Can the data form a closed loop?
Can I get feedback data?
Think about how we learned to read when we were young. At the beginning, the probability of making mistakes was high. Every time we make a mistake, our parents and teachers will tell us where we went wrong and what it should be. It is in this cycle of "action-feedback-correction-action again" that effective learning is achieved.
Similar to the human literacy learning process, machines also need "effective feedback" to achieve continuous learning. If there is no feedback data, there will always be problems in the problematic places, and there will never be progress.
Therefore, effective feedback data is an important part of learning and solves the "learning" problem.
Can the data form a closed loop?
When we can get feedback data, the machine can learn. The next step is to solve the "continuous" problem.
Take the example of literacy. Suppose there are 2 children learning literacy at the same time.
- Child A has a teacher who can counsel at any time to correct mistakes
- Child B can only be instructed by the teacher for 1 days a week
Without a doubt, it must be that child A learns faster and better.
The same is true of the machine. To make the data form a closed loop, it is hoped that it can obtain "real-time" feedback data and get the personal coaching like child A.
The so-called data closed loop is to automatically run the above "action-feedback-correction-re-action" cycle on the machine without human involvement.
A typical example in practical applications is the recommendation system in the e-commerce platform. Just after you watch a pair of basketball shoes, a lot of basketball shoes will be recommended for you in an instant.
Therefore, the data forms a closed loop, which allows the machine to achieve continuous learning and solve the "continuous" problem.
How does face recognition in Google Photos collect feedback data and form a data loop?
Many photo album apps now have a face recognition function that can automatically help you classify photos according to different people. But in actual use, there will be many cases of wrong judgment.
If you don't collect feedback data and let the data form a closed loop, the errors will continue!
Step 1: Ask the user
Some photos are not sharp, or some people’s hairstyles have changed a lot, or some people have removed their makeup...
There are many reasons for the machine to be unsure of its own judgment. This machine is not to blame, and many women who have put on makeup cannot recognize it.
At this time, Google Photos will actively ask the user if 2 avatars are the same person, as shown below:
Step 2: Capability upgrade
If I choose "Same person", when the machine collects user feedback, the first thing is to merge their photos together.
There is one more important thing, the machine learns something:
- The machine knows makeup: it turned out that this man was like this.
- The machine knows how to gain weight: it turns out to be like this after gaining weight.
- The machine knows that it is getting old: when people are old, they will have these changes.
Artificial intelligence tells us that Spider-Man is getting older like this:
There is a very important reason why artificial intelligence has high expectations:
AI can break through the limit of human capabilities through continuous learning. Some even predict that machines will surpass humans in all respects.
In order for the machine to achieve continuous learning, we need to implement 2 conditions:
- Continuously get feedback data to let the machine know where it is good and where it is bad
- If feedback data is added to the closed loop, can the machine learn continuously and improve its capabilities?
There are many issues to consider when assessing whether to use artificial intelligence or not. This series will continue to be updated. Interested friends can add me on WeChat.