Mr. Translator
Published in

Mr. Translator

Direct equivalence problem in the Field of AI

What is the essence behind speech recognition, machine translation, natural language understanding, gene recognition and stock analysis, etc?

The key to the development of artificial intelligence is to find the corresponding “aerodynamics”.
In other words, just as we cannot rely on the flight of simulated birds to get airplanes into the air, but to design aircraft that can fly effectively by mastering the principles of aerodynamics, artificial intelligence does not simply imitate human thoughts and movements, but it is necessary to find “aerodynamics” that is suitable for computers to acquire intelligence.

Today, there are actually three pillars of artificial intelligence, namely, Moore’s Law, big data and mathematical models.
I often like to use the following picture to show the relationship between the above concepts and artificial intelligence (also known as machine intelligence).

Let’s talk about the mathematical model from the point of view of the way of thinking.
We know that if we want a computer that can only solve intelligent problems foolishly, it is necessary to find a corresponding mathematical model.
It is not difficult to do certain problems, such as letting the computer control the production line or even the flight of the moon landing rocket.
However, it is difficult to solve the problem of intelligence that human beings themselves do not understand.
So from 1956 to the early 1970s, scientists all over the world pondered the nature of intelligence and saw whether some methods of cognitive science could be expressed quantitatively, so that computers could be used to solve them. We had been exploring in the dark for more than a decade, and no satisfactory results have been achieved.

What is the core method within the artificial intelligence system? Find the importance of equivalence problem.

The first person to make a breakthrough in a small area of artificial intelligence was Jalinick.
He brilliantly found the equivalent problem of speech recognition and machine translation-the communication problem, and solved the above two intelligent problems by solving the communication problem.
Why is speech recognition (including semantic understanding) a communication problem?
When we talk about recognizing other people’s pronunciation, we think of acoustic and linguistic concepts such as sound waves, vowel consonants, phrases, grammar, context, complete sentences, and so on. In fact, these are all representations, or concepts created by us.

That’s what scientists thought in the past, and they built a bunch of models for each concept.
But it is not easy to find a solution among such a large number of concepts.
So what is the nature of speech recognition?
We might as well go back to where we started and think about it.
By speaking, that is, the expression of oral language, we are trying to make the listener understand what is going on in our minds.
This process can be divided into three stages:
1. The speaker describes the meaning in his mind in words.
2. Words become pronunciation, and sound travels in the air or in telephone lines (including the medium of mobile communication). At the listener, sound waves are turned into signals that can be accepted by our ears and transmitted to the listener’s brain through the auditory nerve.
3. The listener restores the signal to the meaning of the speaker.
The first process mentioned above is actually consistent with the coding process of the information source in the communication. Whether you send an email (or text message) or a photo or video in the communication, the computer must first turn it into a code that can be transmitted in the communication system. This is the coding process.
Although the second process mentioned above is complex, it is actually a process in which information is transmitted in different channels.
The third process is the inverse process of the first process, that is, the decoding process.
Encoding-channel transmission-decoding, this is not a standard communication process?
Since the equivalent problem of speech recognition is the communication problem, the method of solving the communication problem can be used to solve the speech recognition problem.
To see this is what makes Jalinick so special.

Since Shannon put forward the information theory in 1946, the classical communication problems have been solved.
Jalinick uses the known communication model to solve the unknown equivalence problem, that is, the speech recognition problem, which is successful in one fell swoop.
Before Jalinick, the world worked hard for more than a decade to recognize only a few hundred English words, and the error rate was more than 30%. After five or six years of work with dozens of scientists, Jalinick’s way was able to recognize 22000 English words. and reduce the error rate to 10%.

Of course, because the communication model requires a large amount of data to train the model, this method is also known as the “data-driven method”.
Next, one of Jarenick’s men, Peter Brown, found that the problem of machine translation was also a communication problem.
That’s what Brown thinks.
If we want to translate Chinese into English, we usually think that the speaker speaks Chinese.
But Brown looks at the problem from a different point of view. He thinks that when you speak to a foreigner, you say a meaning that can be expressed in English, but when you express this meaning, you choose to use Chinese to do a special coding.
All the translator, the decoding party, has to do is to decode the string of information encoded in Chinese into your (English) meaning.
Looking at the problem in this way, the problem of machine translation has been changed from the understanding of two natural languages to a communication problem.

We can sum up the ideas of Jarenick and Brown in this formula:
Speech recognition ≡ communication problem ≡ machine translation.
(Three of the bars represent mathematical equivalence.
The art of science and engineering often lies in being good at finding known equivalents of unknown complex problems. This should become the methodology of our work. By knowing one method you will know all, we actually know which problems experience can be extended to others after solving a problem. )

As for why Jalinick can think of intelligent problems such as speech recognition is a communication problem, this is mainly related to his previous experience and expertise.
Jalinick has never been a traditional artificial intelligence expert, so in the 1960s and 1970s, that group of people’s ideas did not influence him at all, which allowed him to avoid detours.
But, on the other hand, as a student of Shannon, he was the best communications expert of his time and wrote a textbook on university information theory at that time.

In many cases, success requires all kinds of ripe conditions. In the absence of data, Brown found the right way, but If you have no hand you can’t make a fist. At this time, people who lack wisdom will complain, and people with wisdom will open up a new battlefield.

Since the early 1960s, he has been thinking about using information theory to solve language problems, but in the early 1970s, he came to IBM, to have the opportunity and resources to realize his ideas.

And then, can we find the equivalent of this kind of problem?
Of course, in the 1990s, Professor Salzberg of Johns Hopkins University took a whole new perspective on gene sequencing.
He believes that human genes are just a special book with only four letters A, G, C and T, so any algorithm that recognizes language, such as speech recognition and OCR, can be used for gene sequencing.

With this idea in mind, he left the university and went to TIGR, a research institute that specializes in gene sequencing, and later won the top prize in that field.
Salzberg and Jalinick were colleagues at the university, and his idea was certainly inspired by the latter, but he was the first to see that “gene sequencing and speech recognition, and communication” were equivalent problems each other and used it as the key to solve the problem. that’s what makes him so good.

There are many kinds of equivalence problems in this category.
Briefly speaking they can be summarized as two categories: first category is direct equivalence problem including stock market analysis and prediction; second category is indirect generalized equivalence problem including many image recognition today especially face and medical image recognition problems.

Among the first category of problems mentioned earlier Brown has done a good job and has achieved great success.
When IBM engaged in machine translation due to lack of data so Brown’s findings were not quite good enough and his epoch-making paper didn’t even cited by anyone for more than ten years.

He job-hopping to Renaissance Technologies (i.e. hedge fund firms which has lot of data (their investment returns were highest for nearly 30 years were made). There Brown gained great success.
Later, using Brown’s idea to solve machine translation problem isn’t himself, but rather than Och.
God always fair to everyone, allowing Brown who withdrew from academia to gain money while giving Brown’s honor to later generations of scientists who were young scientist Och.

Although Brown first proposed “the correct framework of machine translation”, he did not become the last inventor because of objective conditions. There are many such people in history, and most of the time, what we see is that these people come out to emphasize their original contributions after the success of others. But Brown disdained to do so, but found his place in new areas and became the real winner.

--

--

--

Interpreter & Dictionary provided by Tencent Cloud & Smart Industries Business Group (CSIG)

Recommended from Medium

The Game of Solving Problems­

Human-centered AI to build a trustful customer experience in retail

Top Use Cases of AI-NLP Insurance Chatbots

How AI Will Reshape Companies, Industries and Nations: An interview with Kai-fu Lee of Sinovation…

AI without Objectives

So, bots you say…

How to Become Super-Rich

Difference between AI vs ML vs DL

Get the Medium app

A button that says 'Download on the App Store', and if clicked it will lead you to the iOS App store
A button that says 'Get it on, Google Play', and if clicked it will lead you to the Google Play store
Chier Hu

Chier Hu

linktr.ee/chierhu linkedin.com/in/chierhu quora.com/profile/Chier-Hu instagram.com/chierhu youtube.com/chierhu twitter.com/chierhu chierhu@qq.com

More from Medium

How AI can Transform Business Intelligence

How NLP will accelerate EU Taxonomy reporting (and make way for a new era in sustainable finance)

Best words for Wordle and Quordle — The Data Science way !

English to Spanish translation with Transformer