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 angle of this issue-black box
List of article series:
Black box is a disadvantage of artificial intelligence
Not all artificial intelligence is a black box. The black box everyone refers to is the hottest and best-effect "deep learning" at the moment.
I wrote beforeI understand deep learning in one article", I gave an example of a faucet, and from that example we can see:The working principle of deep learning is not logic (rules based), but miracles (based on statistics).
A miracle can lead to several results:
- Deep learning can only tell you "what", but cannot tell you "why"
- No one can predict when an error will occur
The picture below shows some of the "low-level mistakes" made by artificial intelligence.
And the most terrible thing is: when we find a problem, we can't prescribe the right problem for the specific problem.
Most of our computer science in the past was rule-based, much like a car, and we knew exactly how the car was assembled, so when we found that the screws were loose, we tightened them, and we replaced them with any parts that were old. Can be done right.
Deep learning is completely different. When we find a problem, we can't do it right. We can only optimize it globally.(Such as filling more data).
Extended reading:
'Depth | Nature: Can we open the "black box" of artificial intelligence?"
'Breaking the artificial intelligence algorithm black box"
What questions aren't suitable for "dependent" AI?
Due to the black box nature of deep learning, not all problems are suitable for solving with deep learning.
When we evaluate which questions are suitable and which questions are not, we can evaluate from 2 perspectives:
- Need explanation
- Error tolerance
Let's first look at AI applications with higher penetration from these 2 perspectives:
| Case | Need explanation | Error tolerance |
|---|---|---|
| Speech Recognition | The user only cares about the effect, and does not care about the principle behind it. | Occasionally some errors do not affect the understanding of the entire sentence. Minor errors are acceptable. |
| Face recognition | Ibid. | Compared with speech recognition, users are less tolerant to errors because they need to re-brush their faces. |
| machine translation | Ibid. | Similar to speech recognition, as long as it is accurate in the big picture, it does not affect the overall understanding. |
Let's look at some specific applications of AI and human integration:
| Case | Need explanation | Error tolerance |
|---|---|---|
| Smart customer service | The user doesn't care whether it is manual service or machine service, as long as it can solve my problem | If the machine customer service cannot understand my intent and cannot help me solve the problem, the user will be very dissatisfied. So when the machine is in doubt, you need to fill up the position manually. |
| Content review | For the content that fails the review, the reason needs to be explained. What is passed need not explain why. | There is a profession called "Jian Huangshi", which is gradually being replaced by a machine, but it has not been completely replaced, because sometimes the machine will be inaccurate. At this time, a manual review will be performed. |
Finally, look at some scenarios that are not suitable for AI landing:
| Case | Need explanation | Error tolerance |
|---|---|---|
| Derivation theorem | Science is absolutely rigorous and must be derived logically, not statistically. | If there is an exception, it cannot be called a theorem, it must be absolutely correct and error-free. |
| write thesis | Artificial intelligence can already write novels, poetry, and prose. But the style of the thesis requires very strict context logic. | Errors are not allowed in the paper, and the logic of the full text must be very clear. Even if there is a logical problem in a detail, the entire paper will be worthless. |
If we put all the cases mentioned above in the quadrant, it is roughly as follows:
Therefore, there are 3 principles in the evaluation:
- The more the solution needs to explain the reason behind it, the less suitable it is to use deep learning
- The lower the tolerance for errors, the less suitable it is to use deep learning
- The above 2 clauses are not absolute judgment standards, but also need to look at commercial value and cost performance. Autonomous driving and medical treatment are counter examples.
Case Study: Medical
The application of artificial intelligence in the medical industry is widely optimistic, because the medical industry has many pain points:
- Insufficient medical resources, especially quality doctors
- The distribution of medical resources is extremely uneven, and many diseases in China can only be treated by Beijing
- In fact, the doctor's misdiagnosis rate is also very high (malignant tumor misdiagnosis rate 40%, organ ectopic misdiagnosis rate 60%)
The current artificial intelligence can help humans make diagnosis and provide treatments.
the strange thing is:Whether in terms of interpretability or tolerance for errors, medical diagnosis is not suitable for artificial intelligence.
But when we use artificial intelligence as an aid, we still rely on humans to make judgments and make decisions. Humans and machines can complement each other very well.
The development of the factory is also a similar path:
- In the beginning, the machine only assisted. Manpower was the most important.
- The degree of mechanization and automation is getting higher and higher, and the role of machines is increasing
- Eventually realized unmanned factory (already implemented)
So from "interpretability" and "error tolerance", you can evaluate which problems are not suitable for "full dependence on artificial intelligence."
But as long as the commercial value is large enough, there is still a solution-humans and machines cooperate with each other to solve problems together. And with the advancement of technology, the demand for manpower is continuously reduced.
Extended reading:
'Artificial intelligence can help diagnose diseases quickly, but it cannot replace clinicians"
'2019 China artificial intelligence medical white paper released (with download)"
My blog is about to sync to the Tencent Cloud + community, and invite everyone to join in: https://cloud.tencent.com/developer/support-plan?invite_code=33ax06m8ysao4
Comments