Do not exaggerate, do not speculate: see the essence of AI through machine learning examples

This article is reproduced from the public number Xinzhiyuan,Original address

The term "artificial intelligence" (AI) was originally born in 1955, but the idea of ​​intelligent machines dates back even further, specifically in ancient Greece, ancient China, and ancient India. Perhaps this is one of the reasons why AI has such a huge effect on our imagination, and as such, there is always so much heated discussion about AI technology.

But AI is not a myth, not a magic, but a technology. Like other technologies, it has gone through decades of research and eventually reached a new level of output.The decline in computing power costs and the existence of massive data are the two major reasons why AI technology has grown.AI and machine learning have become a useful tool in a variety of fields such as astronomy, medical, transportation, and music.

After years of expectation, AI technology is finally gradually becoming practical. But one of the hallmarks of practical technology is that they tend to disappear in the end. We will forget the techniques that were once easy to use, and we should not let AI technology repeat the same mistakes. Any technique requires careful review, especially for AI technologies where the risk of failure is so great.

Let's take a look at how AI and machine learning affect today's technology, because when AI really changes the world, it may be too late to understand these effects.

There is a common psychological phenomenon: if you repeat a word for a certain number of times, the word will eventually lose all meaning and turn into a function word with a shell and no practical meaning.

For many of us,The term "artificial intelligence" has experienced this "meaningless" process a long time ago.In the current technical field, "artificial intelligence" is almost everywhere. From TV to toothbrush, all functions can't wait to be "artificial intelligence", but the meaning of the word itself is increasingly blurred.

This is very bad.

Although the term "artificial intelligence" has been undoubtedly abused, AI technology itself is more developed than ever. It is used for health care and war, helping people make music and books, checking their resumes carefully, judging their credibility, and processing photos taken on your mobile phone. In short, whether you like it or not, artificial intelligence will make decisions that affect your life.

Whether we like it or not, artificial intelligence can make decisions about our daily lives.

But this may be far from the hype and touting of AI by technology companies and advertisers. Take Oran-B's Genius X toothbrush as an example. This is one of the many devices that appeared at CES this year, focusing on the so-called "AI" capability. However, through the words in the press release, this toothbrush only provides very simple feedback, telling you whether to brush your teeth at the right time and place. The smart sensor on the toothbrush can detect the position of the toothbrush in the mouth. Calling this thing "artificial intelligence" is basically nonsense, nothing more.

But when AI technology is not hyped, it often leads to misunderstandings. News reports may exaggerate research results and casually package a vague AI story into a "terminator" level of discovery. This may lead people to be confused about the question "what is AI?"

For ordinary people who are not experts, this can be a tricky topic. People often mistakenly confuse current artificial intelligence with the "artificial intelligence" they are most familiar with: the latter mostly appear to be smarter and more conscious than humans. Computer. Experts refer to this particular form of artificial intelligence as "general artificial intelligence." If we do create something similar, there may be a long way to go in the future. Before that, exaggerating the intelligence level or ability of the AI ​​system is not good for anyone.

Don’t let AI empty, it’s important to understand “what is AI”

So, what exactly is AI?(The picture above is from top to bottom. The clockwise direction is: the model of the movie Metropolis, the AI ​​toothbrush of Oral-B, the automatic delivery robot.)

In fact, it is more appropriate to discuss "machine learning" than to discuss AI.Machine learning is a sub-area of ​​AI that encompasses almost all the methods that have the greatest impact on the world (including "deep learning"). The word has no mystery of "AI", but it is more helpful in explaining the role of technology.

What is the way machine learning works? In the past few years, I have seen dozens of explanations, and the most useful difference is that the word "machine learning" itself: machine learning is to let computers learn on their own. But this has led to a bigger problem.

Let's look at a problem first. Suppose you want to build a program that recognizes cats. You can try to use the old method to program some clear rules, such as "the cat has pointed ears" and "the cat is furry". But what if the program shows a picture of a tiger to the program? It is time-consuming and laborious to formulate all the required rules by programming, and in the process it must involve definitions of various difficult concepts, such as the definition of "spike" and "furry".

SoA better way is to let the machine learn by itself.. Therefore, a large number of cat photos can be provided for the machine, and the system will view the images in a way that is unique to them. At first, it was almost randomly connected to different points, but as the experiment was repeated, the system continued to learn the updated version. In the end, you can more accurately determine which photos are cats and which are not cats.

AI system is easy to "self-learning" and easy to "cut corners"

So far, what we are talking about is something that can be predicted. In fact, you may have seen such an explanation before, but what is important is not the interpretation itself, but the implications it implies. What are the side effects of making a decision-making system adopt such a learning model?

The advantage of this approach is obvious: never actually programming. Of course, there is a lot of work to do, improve the way the system handles the data, and come up with more sensible ways to extract information, but you don't tell the system what to look for. That is to say, the system may discover new patterns that humans may miss or never think of. And because all programs require data, 1 and 0, you can use this system to train a variety of tasks, because the modern world is full of large amounts of data.There is a hammer for "machine learning" in the hands. In the digital world, there are nails that can be nailed.

However, we must also see the shortcomings of this approach.If you haven't explicitly taught a computer, how do you know how it makes a decision? Machine learning systems can't explain their ideas, which means your algorithms may perform well for the wrong reasons. Similarly, computer systems may produce a prejudiced worldview or may only be good at accomplishing a small number of related tasks similar to previously obtained data.

The machine learning system does not have what humans expect.common sense. You can build the world's best performing cat recognition program, but it will never tell you that cats shouldn't drive a motorcycle, or that the cat's common nickname is "Tiddles" instead of "Megalorth the Ondying."

Teaching computers to learn by themselves is a wise shortcut. But like all shortcuts, it involves cutting corners. There is intelligence in the AI ​​system, but it is not organic intelligence and does not follow human rules. For example, the question may be asked: How smart is a book? What expertise has been incorporated into a frying pan?

The Future of AI: Achieving Technological Change of “Smooth and Silent”

So which stage of artificial intelligence are we now in? After the baptism of countless “next major breakthroughs” in news hype, some experts believe that we have reached a stable state. But this is not an obstacle to the advancement of AI technology.In terms of AI research, there is still a broad space for exploration in our existing knowledge, and in terms of products, we have only found the tip of the iceberg of the algorithm.

Venture strategist Benedict Evans likened machine learning to a relational database, an enterprise software that created a lot of wealth in the 90 era in the last century and revolutionized the industry, but when When you read the word "relational database", it may not cause special attention. We are now in the stage of normal and rapid development of artificial intelligence. "Ultimately, almost everything will involve a certain area of ​​(machine learning), no one will make a fuss about it," Evans said.

He is right, but we have not reached this stage yet.

Artificial intelligence and machine learning will continue to be new and often unexplained new areas, both now and in the future, and there are many new unexplored issues. In the future, there will be more and more AI technologies that will promote changes in all aspects of life. One day, you will find that the application of AI has become so common that it will not even attract your attention at all.