<\/a><\/noindex>In his book, Marcus and his co-author Ernest Davis advocate for a different approach. They believe we are still far from creating general AI, but they are confident that it will be achieved sooner or later.<\/p>\n
Why do we need general AI? Specialized versions have already been created and provide significant benefits.<\/b><\/p>\n
That's true, and the benefits will only increase. However, there are many tasks that specialized AI simply cannot solve. For example, understanding everyday speech, or general assistance in a virtual world, or a robot helping with cleaning and cooking. Such tasks are beyond the capabilities of specialized AI. Another interesting practical question: can specialized AI create a safe self-driving car? Experience shows that such AI still has many issues with behavior in abnormal situations, even while driving, which complicates the situation significantly.<\/p>\n
I think we all would like to have an AI that can help us make new large-scale discoveries in medicine. It\u2019s unclear whether current technologies will suffice, as biology is a complex field. You need to be prepared to read a lot of books. Scientists understand the causal relationships in the interactions of networks and molecules, and can develop theories about planets and so on. However, with specialized AI, we cannot create machines capable of such discoveries. With general AI, we would be able to revolutionize science, technology, and medicine. In my opinion, it's very important to continue working on creating general AI.<\/p>\n
It seems that by \"general\" you mean strong AI?<\/b><\/p>\n
When I say \"general,\" I mean that the AI will be able to think on the fly and independently solve new tasks. Unlike, say, Go, where the problem hasn't changed in the last 2000 years.<\/p>\n
General AI must be able to make decisions in both politics and medicine. This is analogous to human capability; any sane person can do a great deal. You take inexperienced students and within a few days get them to work on practically anything, from legal tasks to medical ones. This is because they have a general understanding of the world and can read, allowing them to contribute to a very broad range of activities.<\/p>\n
The relationship between such intelligence and strong intelligence is that weak intelligence probably won't be able to solve general tasks. To create something sufficiently reliable, capable of operating in a constantly changing world, you may need at least to approach general intelligence.<\/p>\n
But right now, we are far from that. AlphaGo can play excellently on a 19\u00d719 board, but it needs to be retrained to play on a rectangular board. Or take an average deep learning system: it can recognize an elephant if it is well lit and the texture of its skin is visible. However, if only the silhouette of the elephant is visible, the system will surely fail to recognize it.<\/p>\n
In your book, you mention that deep learning cannot achieve the capabilities of general AI, as it lacks deep understanding.<\/b><\/p>\n
In cognitive science, there is talk of forming various cognitive models. I sit in a hotel room and understand that there is a closet over there, a bed there, and a television that is unusually mounted. I know all these objects; I'm not just identifying them. I also understand how they are interconnected. I have ideas about how the surrounding world functions. They are not perfect. They may be wrong, but they are quite good. And based on them, I draw many conclusions that guide my everyday actions.<\/p>\n
The other extreme is something like the Atari gaming system created by DeepMind, where it memorized what it needed to do when it saw pixels in certain places on the screen. If you get enough data, it can seem like you have understanding, but in reality, it's very superficial. The proof of this is that if you shift objects by three pixels, the AI performs much worse. Changes confuse it. This is the opposite of deep understanding.<\/p>\n
To address this issue, you suggest going back to classical AI. What advantages should we strive to utilize?<\/b><\/p>\n
There are several advantages. <\/p>\n
Firstly, classical AI is actually a framework for creating cognitive models of the world, based on which conclusions can then be drawn.<\/p>\n
Secondly, classical AI is perfectly compatible with rules. Currently, in the field of deep learning, there is a strange trend where specialists try to avoid rules. They want to do everything using neural networks and avoid anything that looks like classical programming. But there are problems that have been easily solved in this way, and no one paid attention to it. For example, route construction in Google Maps.<\/p>\n
In fact, we need both approaches. Machine learning allows for learning from data well, but it poorly aids in representing the abstraction that a computer program embodies. Classical AI works well with abstractions, but it must be programmed entirely by hand, and there is far too much knowledge out there to program it all. Clearly, we need to combine both approaches.<\/p>\n
This relates to the chapter where you discuss what we can learn from the human mind. Primarily, it concerns the concept based on the aforementioned idea that our consciousness consists of many different systems that operate differently.<\/b><\/p>\n
I think there's another way to explain this: each cognitive system we possess actually solves different tasks. Corresponding parts of AI should be designed to address various tasks that have different characteristics.<\/p>\n
Right now, we are attempting to use some all-in-one technologies to tackle tasks that are fundamentally different from one another. Understanding a sentence is not the same as recognizing an object. Yet, people are trying to apply deep learning in both cases. From a cognitive perspective, these are qualitatively different tasks. I'm just astonished by how little appreciation there is in the deep learning community for classical AI. Why wait for a silver bullet? It is unattainable, and futile searches do not capture the complexity of the task of creating AI.<\/p>\n
You also mention that AI systems are necessary for understanding causal relationships. Do you believe that deep learning, classical AI, or something entirely new will help us in this?<\/b><\/p>\n
This is yet another area where deep learning is not particularly suited. It does not explain the causes of certain events but calculates the probability of an event under given conditions.<\/p>\n
What are we talking about? You look at certain scenarios, and you understand why this happens and what might occur if certain circumstances change. I can look at the stand that holds the television and imagine that if I cut one leg off it, the stand will tip over and the television will fall. This is a causal relationship.<\/p>\n
Classic AI gives us certain tools for this purpose. It can, for example, illustrate what support is and what failure means. However, I won\u2019t overpraise it. The problem is that classical AI mostly depends on the completeness of information about what is happening, and I reached my conclusion just by looking at the stand. Somehow, I can generalize and imagine parts of the stand that are not visible to me. We currently lack tools to implement this property.<\/p>\n
You also mention that people possess innate knowledge. How can this be implemented in AI?<\/b><\/p>\n
At the moment of birth, our brain already represents a carefully thought-out system. It is not fixed; nature created the first rough draft. Then learning helps us revise this draft throughout our lives.<\/p>\n
The rough draft of the brain already possesses certain capabilities. A newborn mountain goat can navigate down a slope accurately just a few hours after birth. It is clear that it already has an understanding of three-dimensional space, its body, and the relationship between them. A rather complex system.<\/p>\n
Partly for this reason, I believe we need hybrids. It is hard to imagine creating a robot that functions well in the world without similar knowledge; starting from a blank slate and learning from prolonged, vast experience seems daunting.<\/p>\n
As for humans, our innate knowledge derives from our genome, which has evolved over a long period. With AI systems, we will have to take a different approach. Part of it might involve rules for constructing our algorithms. Partly, it can be rules for creating data structures manipulated by these algorithms. And partially, there could be knowledge that we directly embed into machines.<\/p>\n
Interestingly, in your book, you lead towards the idea of trust and building trust-based systems. Why did you choose this particular criterion?<\/b><\/p>\n
I believe that today this all represents a game of ball. We seem to be living in a strange moment in history, largely trusting software that is untrustworthy. I think the anxieties we have today will not be eternal. In a hundred years, AI will justify our trust, or perhaps even sooner.<\/p>\n
But today, AI is dangerous. Not in the way that Elon Musk fears, but in that hiring interview systems discriminate against women, regardless of what programmers do, because their tools are too simplistic.<\/p>\n
I wish we had a higher-quality AI. I don't want to see an 'AI winter' when people realize that AI doesn't work and is simply dangerous, and they don\u2019t want to fix it.<\/p>\n
In a sense, your book does seem very optimistic. You suggest that it\u2019s possible to build trustworthy AI. We just need to look in a different direction.<\/b><\/p>\n
Indeed, the book is very pessimistic in the short term and highly optimistic in the long term. We believe that all the problems we\u2019ve described can be solved if we take a broader perspective on what the right answers should be. And we think that if this happens, the world will be a better place.habr.com<\/a><\/p>","protected":false,"gt_translate_keys":[{"key":"rendered","format":"html"}]},"excerpt":{"rendered":"\u042d\u0442\u043e\u0442 \u0442\u0435\u043a\u0441\u0442 \u2014 \u043d\u0435 \u0440\u0435\u0437\u0443\u043b\u044c\u0442\u0430\u0442 \u043d\u0430\u0443\u0447\u043d\u043e\u0433\u043e \u0438\u0441\u0441\u043b\u0435\u0434\u043e\u0432\u0430\u043d\u0438\u044f, \u0430 \u043e\u0434\u043d\u043e \u0438\u0437 \u043c\u043d\u043e\u0433\u0438\u0445 \u043c\u043d\u0435\u043d\u0438\u0439 \u043e\u0442\u043d\u043e\u0441\u0438\u0442\u0435\u043b\u044c\u043d\u043e \u043d\u0430\u0448\u0435\u0433\u043e \u0431\u043b\u0438\u0436\u0430\u0439\u0448\u0435\u0433\u043e \u0442\u0435\u0445\u043d\u043e\u043b\u043e\u0433\u0438\u0447\u0435\u0441\u043a\u043e\u0433\u043e \u0440\u0430\u0437\u0432\u0438\u0442\u0438\u044f. \u0418 \u0437\u0430\u043e\u0434\u043d\u043e \u043f\u0440\u0438\u0433\u043b\u0430\u0448\u0435\u043d\u0438\u0435 \u043a \u0434\u0438\u0441\u043a\u0443\u0441\u0441\u0438\u0438. \u0413\u0430\u0440\u0438 \u041c\u0430\u0440\u043a\u0443\u0441, \u043f\u0440\u043e\u0444\u0435\u0441\u0441\u043e\u0440 \u041d\u044c\u044e-\u0419\u043e\u0440\u043a\u0441\u043a\u043e\u0433\u043e \u0443\u043d\u0438\u0432\u0435\u0440\u0441\u0438\u0442\u0435\u0442\u0430, \u0443\u0432\u0435\u0440\u0435\u043d, \u0447\u0442\u043e \u0433\u043b\u0443\u0431\u043e\u043a\u043e\u0435 \u043e\u0431\u0443\u0447\u0435\u043d\u0438\u0435 \u0438\u0433\u0440\u0430\u0435\u0442 \u0432\u0430\u0436\u043d\u0443\u044e \u0440\u043e\u043b\u044c \u0432 \u0440\u0430\u0437\u0432\u0438\u0442\u0438\u0438 \u0418\u0418. \u041d\u043e \u043e\u043d \u0442\u0430\u043a\u0436\u0435 \u0441\u0447\u0438\u0442\u0430\u0435\u0442, \u0447\u0442\u043e \u0438\u0437\u0431\u044b\u0442\u043e\u0447\u043d\u043e\u0435 \u0443\u0432\u043b\u0435\u0447\u0435\u043d\u0438\u0435 \u044d\u0442\u043e\u0439 \u043c\u0435\u0442\u043e\u0434\u0438\u043a\u043e\u0439 \u043c\u043e\u0436\u0435\u0442 \u043f\u0440\u0438\u0432\u0435\u0441\u0442\u0438 \u043a \u0435\u0451 \u0434\u0438\u0441\u043a\u0440\u0435\u0434\u0438\u0442\u0430\u0446\u0438\u0438. \u0412 \u0441\u0432\u043e\u0435\u0439 \u043a\u043d\u0438\u0433\u0435 Rebooting […]<\/p>\n","protected":false,"gt_translate_keys":[{"key":"rendered","format":"html"}]},"author":1,"featured_media":29302,"comment_status":"open","ping_status":"open","sticky":false,"template":"","format":"standard","meta":{"footnotes":""},"categories":[702],"tags":[],"class_list":["post-39055","post","type-post","status-publish","format-standard","has-post-thumbnail","hentry","category-news"],"aioseo_notices":[],"aioseo_head":"\n\t\t\n\t
![\"We](\"\/wp-content\/uploads\/2019\/10\/dccacd1c137f2ec973edd6dc5d4c42f3.png\")