2019-12-12 20:01:00
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 - Black Box
Series of articles list:
"Black Technology" will be used for the job, discuss how artificial intelligence landing
Black Box is a shortcoming of artificial intelligence
Not all AI is a black box, we say that the black box mainly refers to the currently most popular, the effect is the best "deep learning."
Before I wrote " " 65 PDF "allows PM comprehensive understanding of deep learning " in, for example through a faucet, you can see from the example: the working principle of the depth of learning is not about logic (rule-based), but vigorously miracle (based on statistics) .
Vigorously miracle cause several results:
- Deep learning can only tell you "what", but can not tell you "why"
- No one can predict the circumstances under which the error occurs
The picture below will show some artificial intelligence committed "a stupid mistake."
The most frightening thing is: When we find problems, not to remedy specific problems.
Most of our past in computer science is rule-based, much like a car, we clearly know how this car is assembled, it is found that a screw loose to tight lemon, which is a part of aging on the exchange. You can do the right remedy.
The depth of learning is completely different, when we find problems, can not do the right remedy, can only global optimization (such as filling more data).
What issues do not fit "dependent" AI?
Since the black box depth study of the characteristics, not all issues are suitable for deep learning to solve.
What are the problems we assess fit, what issues do not fit, it can be assessed from two perspectives:
- Need to explain
- Error tolerance
We start to look at these two angles higher penetration AI applications:
We look at some specific applications of AI and human combination:
Finally, look at some of the AI is not suitable for landing scenario:
If the case we mentioned above all on quadrants, as follows:
So, there are three principles at the time of the assessment:
- Solutions need to explain the reasons behind it, the less suitable for deep learning
- The lower the tolerance for error, the less suitable for deep learning
- The above two is not an absolute criterion, need to look at the business value and cost-effective, autopilot and health is the counter-example.
Case Study: Medical
It is widely optimistic in artificial intelligence applications in the medical industry, the medical industry because there are many pain points:
- Lack of medical resources, especially high-quality doctor
- Allocation of medical resources is extremely uneven, many Chinese disease can be cured only Beijing
- In fact, doctors misdiagnosis rate is also high (cancer misdiagnosis rate of 40%, 60% misdiagnosis rate of ectopic organ)
The current artificial intelligence may have helped humans to make a diagnosis and provide treatment.
The strange thing is: either from the interpretability or from the error tolerance in terms of medical diagnosis are not suitable for artificial intelligence.
But we will artificial intelligence as an aid, ultimately rely on human judgment and decide to do. Humans and machines can form a good complement.
Development of the factory is also a similar path:
- Only the beginning of a secondary machine, manpower is the most important
- The degree of mechanization and automation have become increasingly demanding, increasing the role of the machine
- Ultimately nobody plant (already achieved)
So from the "interpretability" and "error tolerance" can be evaluated out what the problem is not suitable, "totally dependent on artificial intelligence."
But as long as the commercial value is large enough, there are solutions - humans and machines complement each other to jointly solve the problem. And as technology advances, declining demand for manpower.