This text is not the result of a scientific study, but one of many opinions regarding our next technological development. And at the same time an invitation to discussion.
Gary Markus, a professor at New York University, is convinced that deep learning plays an important role in the development of AI. But he also believes that excessive enthusiasm for this technique can lead to its discredit.
In his book Rebooting AI: Building artificial intelligence we can trust Marcus, a neuroscientist by training who has built a career on cutting-edge AI research, addresses technical and ethical issues. From a technology standpoint, deep learning can successfully mimic the perceptual tasks that our brains perform, such as recognizing images or speech. But for other tasks, such as understanding conversations or determining cause-and-effect relationships, deep learning is not suitable. To create more advanced intelligent machines capable of solving a wider range of problems - often called artificial general intelligence - deep learning must be combined with other techniques.
If an AI system does not truly understand its tasks or the world around it, this can lead to dangerous consequences. Even the smallest unexpected changes in the environment of the system can lead to its erroneous behavior. There have already been many such examples: determiners of inappropriate expressions that are easy to deceive; job search systems that constantly discriminate; self-driving cars that crash and sometimes kill the driver or pedestrian. Creating artificial general intelligence is not just an interesting research problem, it has many very practical applications.
In their book, Marcus and his co-author Ernest Davis put forward a different path. They believe that we are still far from creating a general AI, but they are confident that sooner or later it will be possible to create it.
Why do we need general AI? Specialized versions have already been created and bring a lot of benefits.
True, and there will be even more benefits. But there are many problems that specialized AI simply cannot solve. For example, understanding ordinary speech, or general help in the virtual world, or a robot helping with cleaning and cooking. Such tasks are beyond the capabilities of specialized AI. Another interesting practical question is: can you use specialized AI to create a safe self-driving car? Experience shows that such AI still has many problems with behavior in anomalous situations, even when driving, which complicates the situation quite a lot.
I think we all would like to have AI that can help us make new large-scale discoveries in medicine. It is unclear whether current technologies will be suitable for this, because biology is a complex field. You have to be prepared to read a lot of books. Scientists understand the cause-and-effect relationships in the interaction of networks and molecules, they can develop theories about planets, and so on. However, with specialized AI, we cannot create machines capable of such discoveries. And with general AI, we could revolutionize science, technology, and medicine. In my opinion, it is very important to continue working on the creation of general AI.
It seems that by "general" you mean strong AI?
By "general" I mean that the AI ββwill be able to think on the fly and independently solve new problems. Unlike, say, Go, in which the problem has not changed for the last 2000 years.
General AI should be able to make decisions in both politics and medicine. This is analogous to human ability; any sane person can do a lot. You take inexperienced students and in a few days you have them working on almost anything from a legal problem to a medical one. This is due to the fact that they have a general understanding of the world and can read, and therefore can contribute to a very wide range of activities.
The relationship between such an intellect and a strong one is that a weak intellect will probably not be able to solve common problems. To create something robust enough to handle an ever-changing world, you may need to at least get close to general intelligence.
But now we are very far from that. AlphaGo can play fine on a 19x19 board, but needs to be retrained to play on a rectangular board. Or take the average deep learning system: it can recognize an elephant if it is well lit and the texture of its skin is visible. And if only the silhouette of an elephant is visible, the system will probably not be able to recognize it.
In your book, you mention that deep learning is not capable of reaching the capabilities of general AI because it is not capable of deep understanding.
In cognitive science, they talk about the formation of various cognitive models. I am sitting in a hotel room and I understand that there is a closet, there is a bed, there is a TV that is unusually suspended. I know all these things, I don't just identify them. I also understand how they are interconnected with each other. I have ideas about the functioning of the surrounding world. They are not perfect. They may be wrong, but they are quite good. And based on them, I make a lot of inferences that become a guide for my daily actions.
At the other extreme, it's something like DeepMind's Atari game system, where it remembers what it needs to do when it sees pixels in certain places on the screen. If you get enough data, it may seem that you have an understanding, but in reality it is very superficial. The proof of this is that if you move objects by three pixels, then AI plays much worse. Changes confuse him. This is the opposite of deep understanding.
To solve this problem, you propose to return to classical AI. What advantages should we try to use?
There are several advantages.
First, classical AI is actually a framework for creating cognitive models of the world, on the basis of which you can then draw conclusions.
Secondly, classical AI is perfectly compatible with the rules. There is a strange trend in deep learning right now where people try to avoid the rules. They want to do everything on neural networks and not do anything that looks like classical programming. But there are tasks that were calmly solved in this way, and no one paid attention to this. For example, building routes in Google Maps.
In fact, we need both approaches. Machine learning is good at learning from data, but very poor at mapping the abstraction that a computer program is. Classical AI works well with abstractions, but it has to be programmed entirely by hand, and there is too much knowledge in the world to program them all. Obviously, we need to combine both approaches.
This is related to the chapter in which you talk about what we can learn from the human mind. And first of all, about the concept, based on the idea mentioned above, that our consciousness consists of many different systems that work in different ways.
I think there's another way to explain it: every cognitive system we have really does a different job. Similar parts of AI must be designed to solve different problems that have different characteristics.
Now we are trying to use some all-in-one technology to solve problems that are fundamentally different from each other. Understanding a sentence is not the same as recognizing an object. But people are trying to use deep learning in both cases. From a cognitive point of view, these are qualitatively different tasks. I'm just amazed at how little classical AI is appreciated in the deep learning community. Why wait for the silver bullet? It is unattainable, and fruitless searches do not allow us to comprehend the full complexity of the task of creating AI.
You also mention that AI systems are needed to understand cause and effect relationships. Do you think that deep learning, classical AI, or something completely new will help us with this?
This is another area where deep learning is not well suited. It does not explain the causes of some events, but calculates the probability of an event under given conditions.
What are we talking about? You watch some scenarios and you understand why this is happening and what can happen if some circumstances change. I can look at the TV stand and imagine that if I cut off one of its legs, the stand will flip over and the TV will fall. This is a causal relationship.
Classical AI gives us some tools for this. He can imagine, for example, what is a support and what is a fall. But I won't exaggerate. The problem is that classical AI is largely dependent on the completeness of the information about what is happening, and I figured it out just by looking at the coaster. I can somehow generalize, represent parts of the stand that I can't see. We don't yet have the tools to implement this property.
You also talk about people having innate knowledge. How can this be implemented in AI?
At the time of birth, our brain is already a very elaborate system. It is not fixed, nature has created the first, rough draft. And then learning helps us revise that draft throughout our lives.
The rough draft of the brain already has certain capabilities. A newborn mountain goat is able to accurately descend the mountainside in a few hours. Obviously, he already has an understanding of three-dimensional space, his body and the relationship between them. A very complex system.
This is partly why I think we need hybrids. It's hard to imagine how a robot can be made to function well in a world without similar knowledge, where to start, instead of starting from scratch and learning from a long, vast experience.
As far as humans are concerned, our innate knowledge comes from our genome, which has evolved over a long time. And with AI systems, we will have to go the other way. In part, these may be the rules for constructing our algorithms. In part, it may be the rules for creating data structures that these algorithms manipulate. And partly it can be knowledge that we will directly invest in machines.
It is interesting that in the book you bring to the idea of ββtrust and the creation of trust systems. Why did you choose this criterion?
I believe that today all this is a ball game. It seems to me that we are living in a strange moment in history, trusting a lot of software that is not trustworthy. I think that the worries that we have today will not last forever. In a hundred years, AI will justify our trust, and maybe sooner.
But today AI is dangerous. Not in the sense that Elon Musk fears, but that job interview systems discriminate against women, no matter what programmers do, because their tools are too simple.
I wish we had better AI. I don't want to start an "AI winter" when people realize that AI doesn't work and is just dangerous, and they don't want to fix it.
In some ways, your book does seem very optimistic. You assume that it is possible to build a trustworthy AI. We just need to look in a different direction.
True, the book is very pessimistic in the short term and very optimistic in the long term. We believe that all the problems we have described can be solved by looking more broadly at what the correct answers should be. And we think that if that happens, the world will be a better place.
Source: habr.com