Boris Katz has built a career helping the machines to learn the language. He believes that the current AI technologies are insufficient to make Siri or Alexa really smart. Siri, Alexa, Google Home technologies that analyze language, often find their application in everyday life. But Boris Katz, chief research officer of MIT, is not impressive. Over the past 40 years, he played a key role in the linguistic abilities of the machines.
In the 1980s he developed a system of START, are able to respond to the formulated natural language queries. Ideas used at the START, helped Watson win Jeopardy! and laid the groundwork for today’s chatbots.
But now Katz is concerned that this area relies on the idea that for many years, but these ideas did not bring machine intelligence to real. MIT Technolody Review interviewed Boris Katz. Let’s find out where you need to send the studyto become smarter.
How to make artificial intelligence is really smart
Where did your story begin teaching computers to use language?
I first encountered computers in the 1960-ies as a student of the Moscow University. Car I used was called BESM-4. To contact her, you could only use octal code. My first computer project involved training of computer reading, understanding and solving mathematical problems.
Then I developed a computer program, or writing poetry. I still remember standing in the engine room waiting for the next poems generated by the machine. I was stunned by the beauty of the verses; they seemed to be created by an intelligent being. Then and there I realized that I wanted the rest of my life to work on the creation of intelligent machines and finding ways to communicate with them.
What do you think about Siri, Alexa, and other personal assistants?
It’s funny to say this, because, on the one hand, we are very proud of the incredible progress in every pocket there’s something we helped to create many years ago and it is wonderful.
But on the other hand, these programs are incredibly stupid. So a sense of pride punctuated by a sense of shame. You launch something that people consider reasonable, but it’s not even close to that.
Thanks to machine learning, in artificial intelligence there has been a significant progress. Doesn’t that make machines better at understanding the language?
On the one hand, there is this dramatic progress, but on the other part of this exaggerated progress. If you look at the achievements of machine learning, all ideas were 20-25 years ago. Just the engineers in the end did a great job and brought these ideas to life. Whatever this technology is great may be, it does not solve the problem of this understanding of real intelligence.
At a very high level of modern methods — statistical methods such as machine learning and deep learning, is very good for finding patterns. And since people usually produce the same deals most of the time, they are very easy to find in the language.
Look at predictive text input. The machine knows better than you what you have to say. You can call it reasonable, but really she just thinks of words and numbers. As we constantly say the same thing, it’s very easy to create systems that capture regularities and behave as if they are reasonable. This fictitious character of much of modern progress.
How about “dangerous” tool for generating language, recently submitted OpenAI?
These examples are really impressive, but I don’t quite understand what they teach us. OpenAI language model was trained on 8 million web pages to predict the next word given all previous words in a text (on the same topic). This huge amount of training, of course, provides local coherence (syntactic and even semantic) text.
Why do you think artificial intelligence is moving in the wrong direction?
In processing language, as in other areas, progress has been made in the training of models on huge amounts of data — millions of sentences. But the human brain cannot learn a language using this paradigm. We don’t leave our children with an encyclopedia in the crib, expecting that they would learn the language.
When we see something, we describe this language; when we hear someone speaks, we present how the described objects and events appear 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 simultaneously learn the language. Perhaps by studying these methods in isolation, we made the problem harder, not easier?
Why is common sense important?
For example, your robot helps you collect things, you say: “This book will not fit in the red box, because it’s too small”. Of course, you want the robot realized that the red box is too small and you could continue a meaningful conversation. But if you tell the robot: “This book will not fit in the red box, because it’s too big”, the robot has to guess that this book is very large, and not a box.
Understanding what the essence of the conversation is a reference, it is very important and people perform this task every day. However, as you can see from these and other examples, it often relies on a deep understanding of the world, which is not currently available for our machines: understanding common sense and intuitive physics, an understanding of the beliefs and intentions of others, the ability to visualize and reason about cause and consequence, and much more.
You are trying to teach machine language, using simulated physical worlds. Why?
I have not seen a child whose parents put the encyclopedia in the crib and say, “Go and learn”. But so do our computers today. I don’t think these systems will be to learn as much as we want, or understand the world the way we want.
In the case of children, they immediately receive tactile sensations from the world. Then the baby starts to see the world and absorb the events and properties of objects. Then the child hears the language input. And the only way the magic happens understanding.
Which approach is best?
One way forward is to obtain a deeper understanding of human intelligence, and then use this understanding to create intelligent machines. The study of AI should be based on the ideas of developmental psychology, cognitive science and neurobiology, and AI should reflect what is already known about how people learn and understand the world.
Real progress will begin only when the scientists come out of their offices and begin to chat with people in other areas. Together we will approach the understanding of intelligence and figuring out how to reproduce it in intelligent machines that can speak, see and act in our physical world.
I agree with Boris? Tell us in our chat in Telegram.