Isn't there too much noise around machine learning? This question, once asked on Quora , is answered by Scott Aaronson, a theoretical computer scientist at the Massachusetts Institute of Technology (Cambridge, Massachusetts, USA), who will soon be working at the University of Texas at Austin.
Topics for discussion are very different, and their value too. "The ideological significance of toenails", for example. Few people talk about this, and, perhaps, there is no need to talk more. Or, semi-exponential functions. Little is said about them, although they should. At the same time, talk about football, the uncivilized way of sitting with legs spread on public transport, or celebrity dresses at the Oscar ceremony does not stop. Although it makes little sense. Finally, there are topics that are talked about a lot, and at first glance, these are really worth talking about: World War II, global warming, black holes or machine learning. It is difficult to say how important they are, since it is very difficult to extract the essence from a pile of informational garbage. In addition, to truly understand a science-intensive topic, you need to have some preparation.deep learning ) and has an impact on society (the same transition to self-driving cars).
I would love to work with machine learning if I wasn't doing quantum computing. In fact, I started to look at artificial intelligence while studying at Cornell University (Ithaca, New York, USA) with Bart Selman (Bart Selman), and then as a graduate student at Berkeley (University of California, Berkeley, California, USA) with Mike Jordan before moving on to quantum computing. I decided that in this area my "comparative advantage" is much greater. The progress in machine learning over the last ten years is the progress that has led to things like IBM Watson, AlphaGoand so on - seems really amazing to me. On the other hand, at least according to the Machine Learning researchers I know, this very progress that everyone is talking about does not include any major new conceptual breakthroughs. It has more to do with refining the algorithms that existed in the 70s and 80s and implementing those algorithms on computers that are orders of magnitude faster and training them with much more data. On the one hand, the fact that ideas that have been maturing for several decades (for example, the backpropagation method - Backpropagation and its variations) were able to show such a high result when scaling is encouraging. After all, we can achieve even greater results of AI,
On the other hand, it is also a reminder that if we want to know what will be important in decades, we need to look for today's analogues of the scientific ideas of the past, the same backpropagation method. Perhaps these ideas, developed by a group of eccentric scientists, are too strange, new and unproven to attract investors or be described in glossy magazines. Ideas that are underestimated due to insufficient development of technology. In conclusion, I want to say that I am much more interested in the very essence of machine learning than all this excitement around it. After all, back in the fifties, almost everyone knew that computers would be important in the future, and they were right. But in many ways, the scientists were wrong. Thus, the complexity of creating humanoid robots was greatly underestimated, people failed to foresee the arrival of personal computers or the Internet. I have no doubt that thirty years from now, people will agree with us about the importance of the role of machine learning. But perhaps they will condemn us for ignoring some areas of Machine Learning, or, on the contrary, laugh at us for tackling issues that have not found further application? I do not know the answers to these questions, but I would like to know them.
GO TO FULL VERSION