Boris Katz built a career helping machinesmaster the language. He believes that the current AI technology is not enough to make Siri or Alexa really smart. Siri, Alexa, Google Home - technologies that analyze the language, more and more often find their application in everyday life. But Boris Katz, MIT's chief research officer, is not impressive. Over the past 40 years, he has made a key contribution to the linguistic abilities of machines.
In the 1980s, he developed the START system, capable ofrespond to natural language requests. The ideas used in START helped Watson win at Jeopardy! and laid the foundation for today's chatbot.
But now Katz is concerned that this areaIt is based on ideas that have been for many years, and these ideas do not bring the machine intelligence closer to the real one. MIT Technolody Review interviewed Boris Katz. Let's find out where you need to direct research to make cars smarter.
How to make artificial intelligence really smart
How did your computer learning history begin?
I first encountered computers in the 1960syears, as a student at Moscow University. The car I used was called BESM-4. Only octal code could be used to communicate with it. My first computer project included teaching the computer to read, understand, and solve mathematical problems.
Then I developed a computer programwriting poems. I still remember how I stood in the machine room, waiting for the next poem created by the machine. I was stunned by the beauty of the poems; it seemed that they were created by an intelligent being. Both then and there I realized that I wanted for the rest of my life to work on creating intelligent machines and finding ways to communicate with them.
What do you think of Siri, Alexa and other personal assistants?
It's funny to talk about it, because, on the one hand, we are very proud of this incredible progress - everyone in his pocket has something that we helped to create many, many years ago, and this is wonderful.
But on the other hand, these programs are incredibly stupid. So a sense of pride is punctuated by a sense of shame. You run something that people think is reasonable, but it’s not even close.
Thanks to machine learning, significant progress has been made in artificial intelligence. Doesn't that make cars better at understanding languages?
On the one hand, there is this dramaticprogress, but on the other hand, part of this progress is bloated. If you look at the achievements of machine learning, all ideas appeared 20-25 years ago. Just as a result, engineers did a great job and brought these ideas to life. No matter how great this technology is, it will not solve the problem of real understanding - true intelligence.
At a very high level, modern methods -Statistical methods, such as machine learning and deep learning, are very good for finding patterns. And since people usually produce the same sentences most of the time, they are very easy to find in the language.
Look at the predictive text input. The machine knows better what you are going to say. You may call it sensible, but in fact she simply counts words and numbers. Since we are constantly saying the same thing, it is very easy to create systems that catch patterns and behave as if they are reasonable. This is the fictitious nature of most modern progress.
How about a “dangerous” tool for generating a language recently introduced by OpenAI?
These examples are really impressive, but I don’tI completely understand what they teach us. The OpenAI language model was trained on 8 million web pages to predict the next word, taking into account all the previous words in a specific text (on the same topic). This vast amount of learning certainly ensures local consistency (syntactic and even semantic) of the text.
Why do you think artificial intelligence is moving in the wrong direction?
In the processing of language, as in other areas, wasProgress has been made in training models on huge amounts of data — millions of sentences. But the human brain cannot learn a language using such a paradigm. We do not leave our children with an encyclopedia in the crib, expecting them to learn the language.
When we see something, we describe it in language;when we hear someone say something, we imagine how the described objects and events look in the world. People live in a physical environment filled with visual, tactile and linguistic sensory data, and the redundant and complementary nature of these inputs allows children to make sense of the world and at the same time learn a language. Perhaps by studying these methods separately, we made the problem more difficult, and not easier?
Why is common sense important?
Let's say your robot helps you collect things, andyou say to him: "This book will not fit in the red box, because it is too small." Of course, you want the robot to understand that the red box is too small and you can continue the meaningful conversation. But if you tell the robot: “This book will not fit in the red box, because it is too big,” the robot must guess that this book is very big, and not the box.
Understanding what the essence of the conversation goessending is very important, and people perform this task every day. However, as you could see from these and other examples, it often relies on a deep understanding of the world that is currently unavailable to our machines: understanding common sense and intuitive physics, understanding the beliefs and intentions of others, the ability to visualize and reason about and consequence, and much more.
You are trying to teach machines the language using simulated physical worlds. Why?
I have not seen a child whose parents putEncyclopedia in the crib and say: "Go learn." But so do our computers today. I do not think that these systems will learn how we want, or understand the world the way we want.
In the case of children, they immediately receivetactile sensations of the world. Then babies begin to see the world and absorb the events and properties of objects. Then the child hears language input. And this is the only way the magic of understanding is created.
Which is the best approach?
One way to move forward is togaining a deeper understanding of human intelligence, and then using that understanding to create intelligent machines. AI research should be based on the ideas of developmental psychology, cognitive science and neurobiology, and AI models should reflect what is already known about how people study and understand the world.
Real progress will only begin whenscientists will come out of their offices and start communicating with people in other areas. Together, we will get closer to understanding the intellect and figuring out how to reproduce it in intelligent machines that can talk, see, and act in our physical world.
Do you agree with Boris? Tell us in our chat in Telegram.