This text is not the result of scientific research, but one of many opinions regarding our immediate technological development. And at the same time an invitation to the discussion.
Gary Marcus, 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 neurologist by training who built his career on cutting-edge AI research, addresses technical and ethical issues. From a technology point of view, deep learning can successfully mimic the solution to the perception tasks that our brain performs: for example, recognition of images or speech. But to solve other problems, such as understanding conversations or determining causal relationships, deep learning is not good. To create more advanced intelligent machines that can solve a wider range of tasks - they are often called general artificial intelligence - deep learning must be combined with other techniques.
If the AI system does not really 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: determinants of inappropriate expressions that are easy to fool; job search systems that constantly discriminate; unmanned vehicles that get into accidents and sometimes kill a driver or a pedestrian. The creation of general artificial intelligence is not just an interesting research problem, it has many completely practical applications.
In his book, Marcus and his co-author Ernest Davis put forward advocate a different path. They believe that we are still far from creating a common AI, but they are sure that sooner or later it will be possible to create it.
Why do we need a common AI? Specialized versions have already been created and bring a lot of benefits.
True, and the benefits will be even greater. But there are many tasks that specialized AI simply cannot solve. For example, understanding ordinary speech, or general help in the virtual world, or a robot that helps with cleaning and cooking. Such tasks are beyond the capabilities of specialized AI. Another interesting practical question: is it possible to create a safe drone car using a specialized AI? Experience shows that such an AI still has many problems with behavior in abnormal situations, even when driving, which greatly complicates the situation.
I think we all would like to get an AI that can help us make new large-scale discoveries in medicine. It is unclear whether current technologies are suitable for this, because biology is a complex field. One must be prepared to read many books. Scientists understand the cause-effect relationships in the interaction of networks and molecules, can develop theories about planets and so on. However, with specialized AI, we cannot create machines capable of such discoveries. And with common AI, we could revolutionize science, technology, and medicine. In my opinion, it is very important to continue working on a common AI.
Looks like “general” do you mean strong AI?
Saying "general" I mean that the AI will be able to think on the fly and independently solve new problems. Unlike, say, Guo, in which the problem has not changed for the past 2000 years.
General AI should be able to make decisions both in politics and in medicine. This is an analogue of human ability; any sane person can do a lot. You take inexperienced students and after a few days force them to work on almost anything, starting with the legal task and ending with the medical one. This is due to the fact that they have a common understanding of the world and are able to 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 non-strong intellect will probably not be able to solve common problems. To create something reliable enough that can work with an ever-changing world, you may need to at least get closer to common intelligence.
But now we are very far from this. AlphaGo can play perfectly 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 its skin texture 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 able to achieve 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’s a closet, there’s a bed, there’s a TV that is unusually suspended. I know all these items, I do not just identify them. I also understand how they are interconnected. I have ideas about the functioning of the world. They are not perfect. They may be wrong, but they are very good. And based on them, I make a lot of conclusions, which become a guide for my daily actions.
The other extreme is something like the Atari gaming system created by DeepMind, in which he remembered what he needed to do when he saw pixels in certain places on the screen. If you get enough data, it may seem that you have an understanding, but in fact it is very superficial. Proof of this is that if you move objects by three pixels, then the AI plays much worse. Change is baffling him. This is the opposite of deep understanding.
To solve this problem, you propose to return to the classic AI. What advantages do we need to try to use?
There are several advantages.
First, the classic AI is actually a framework for creating cognitive models of the world, on the basis of which conclusions can then be drawn.
Secondly, classic AI is perfectly compatible with the rules. Now in the field of deep learning there is a strange tendency when specialists try to avoid the rules. They want to do everything on neural networks and not do anything that looks like classic programming. But there are tasks that were calmly solved in this way, and nobody paid attention to it. For example, building routes in Google Maps.
In fact, we need both approaches. Machine learning allows you to learn well from data, but it helps very poorly in displaying the abstraction that a computer program represents. Classic AI works well with abstractions, but it needs to be fully programmed manually, 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 above-mentioned idea that our consciousness consists of many different systems that work in different ways.
I think there is another way to explain this: each cognitive system that we have really solves different problems. Similar parts of AI should be designed to solve various problems that have different characteristics.
Now we are trying to use some all-in-one technologies to solve problems that are fundamentally different from each other. Understanding a sentence is not at all the same as recognizing an object. But people in both cases try to use deep learning. From a cognitive point of view, these are qualitatively different tasks. I am just amazed at how little the community of deep learning experts value classic AI. Why wait for a silver bullet to appear? It is unattainable, and fruitless searches do not allow to comprehend the complexity of the task of creating AI.
You also mention that AI systems are necessary for understanding cause-effect relationships. Do you think that deep learning, classic AI, or something completely new will help us in this?
This is another area for which deep learning is not too suitable. 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 look at certain scenarios, and you understand why this happens and what can happen if some circumstances change. I can look at the stand on which the TV stands, and imagine that if I cut off one of her legs, the stand will turn over and the TV will fall. This is a causal relationship.
Classic AI gives us some tools for this. He can imagine, for example, what support is and what fall is. But I will not praise. The problem is that the classic AI for the most part depends on the completeness of information about what is happening, but I concluded only by looking at the stand. Somehow I can generalize, imagine parts of the stand that are not visible to me. We do not yet have tools to implement this property.
You also say that people have innate knowledge. How can this be implemented in AI?
At the time of birth, our brain is already a very carefully thought out system. It is not fixed, nature created the first, rough draft. And then learning helps us revise this draft throughout our lives.
A rough draft of the brain already has certain capabilities. A newborn mountain goat in a few hours is able to accurately descend the slope of the mountain. Obviously, he already has an understanding of three-dimensional space, his body and the relationship between them. Very complex system.
This is partly why I think we need hybrids. It is hard to imagine how you can create a robot that functions well in the world without similar knowledge, where to start it, instead of starting from scratch and learning from a long, vast experience.
For humans, our innate knowledge comes from our genome, which has evolved over time. And with AI systems we have to go the other way. In part, these may be the rules for constructing our algorithms. In part, these may be rules for creating data structures that these algorithms manipulate. And in part, this may be knowledge that we will directly invest in machines.
Interestingly, in the book you bring to the idea of trust 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 a strange moment in history, largely trusting software that is not trustworthy. I think the inherent anxieties of today will not last forever. In a hundred years, AI will justify our trust, and maybe even earlier.
But today, AI is dangerous. Not in the sense that Elon Musk fears, but in the fact that job interview systems discriminate against women, regardless of what the programmers do, because their tools are too simple.
I would like for us to have a better AI. I don’t want the “winter of artificial intelligence” to begin, when people realize that AI is not working and just dangerous, and they don’t want to fix it.
In a way, your book really seems very optimistic. You suggest that you can build a credible 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 described by us can be solved if we look more broadly at what the correct answers should be. And we think that if this happens, the world will become better.