To cut through some of the drama that risks being amplified by profiles like this, here are Schmidhuber's [1] and Goodfellow's [2] papers, plus the NIPS reviews [3] that recognized the linkage (not equivalence) between the two.
This is the review process working as it should: helping situate work in a longer dialog so that readers can follow a chain of reasoning. Much of the rest of the drama discussed in this article is about allocating the celebrity that manifests when a technique suddenly becomes useful, through some combination of changes to the environment, theoretical improvements and reduction to practice. It's a very human concern, but it's more business (or politics) than science.
Thanks for this, the reviews for this paper were very interesting.
I do wish that they provided the meta-review, however, which summarizes the ultimate decision. Meta-reviews are written by area chairs, which are the final arbiters for whether or not the paper is accepted. Seeing their perspective on this paper would be very interesting.
For what it's worth, I'm certainly glad Schmidhuber is around. Not only for his great accomplishments but also for the humanity he brings to science and scientific research. In history, accomplishments of this kind are overlooked at times as nothing more than footnotes in time completely overlooking the human dynamics that exist in these highly competitive communities. The "drama" is just as valuable as the resulting research. How many remember or even know the spats between Newton and Leibniz? Tesla and just about everyone in his time (Einstein, Marconi, etc)? He's right in bringing attention to all those that contribute and to demand proper recognition.
Moreover, I also appreciate his focus of AI outside of the marketing/advertising world. Google - just as an example - employees more PhDs than any other public or private organization. All that brain power for the sake of creating a better human trap - aka adtech. Jürgen's most important plea is to use all this AI for real and tangible human progress. Nothing wrong with that in my book.
A related article by the same author (Ashlee Vance) two days later:
Apple and Its Rivals Bet Their Futures on These Men’s Dreams – An oral history of artificial intelligence, as told by its godfathers, gadflies, and Justin Trudeau.
When LeCun only speculates on when the next advances (5-15 years) will come and thinks that it will "take decades" to build systems on what we have... I understand why "His (Schmidhuber's) peers wish he’d just shut up."
Most software excels in a single domain: driving, flying, recording, parsing, etc.
I think AI being smarter than us in a single domain like investing is feasible. But to be smarter than us in all domains, even parenting, love, intuition, etc. sounds infinitely more challenging.
I think AI is usually thought of in terms of solving computationally-heavy or data-heavy problems. But to be "smarter than us", it would need to be better at emotional, conversational and relational problems as well.
Investing is not even remotely a single domain, unless you constrain it down to something fairly trivial like making minority investments in publicly traded US common stocks where being smarter doesn't really count for much. There is so much trading activity now that the markets have become quite efficient. In order to consistently generate alpha as an investor like Warren Buffet you need access to material nonpublic information (proprietary research), or some form of structural advantage. Those things require a huge range of human skills, not just raw smarts.
Lets use IQ as definition of intelligence, whether you agree that an IQ test is a good test for intelligence or not. I actually think it's doable to make a neural net that learns IQ tests, that will become better then the average human, then, by definition, is the "AI" more intelligent ?
The issue with this is the different meanings of intelligence, pimarily 'hard is easy' (chess, go, IQ) vs 'easy is hard' (walking, balance, seeing) test. Once AI robots can win the Champions league at football (soccer) I'll be more impressed.
I can make a robot that will at least tie at Champions league. It would be pretty big, though - exactly the size of the net opening. Obviously, it would play goalie...
Shoot, I'd be impressed if someone could build a robot that lasts an entire game and is agile and light enough to guarantee there wouldn't be multiple broken bones on the opposing team every game.
Also, "Deep Learning in Neural Networks: An Overview" ("88 pages, 888 references"!)
> Abstract. In recent years, deep artificial neural networks (including recurrent ones) have won numerous contests in pattern recognition and machine learning. This historical survey compactly summarises relevant work, much of it from the previous millennium. Shallow and deep learners are distinguished by the depth of their credit assignment paths, which are chains of possibly learnable, causal links between actions and effects. I review deep supervised learning (also recapitulating the history of backpropagation), unsupervised learning, reinforcement learning & evolutionary computation, and indirect search for short programs encoding deep and large networks.
I'd never heard of him, but certainly I'd heard of his LSTM network. Mostly I've learned about the well know members of the AI community through Andrew Ng's interviews with them in his coursera courses. Jürgen Schmidhuber is a notable absentee (but I haven't done all the courses yet).
Plus the company he works for Nnaisense was the NIPS 2017 winner for the 'Reinforcement learning with musculoskeletal models' mentioned yesterday here on HN [0].
Even worse than having your peers crap on you, is the author stating that you wish these crappy remarks don't define your profile (like they did previous profiles), and then having exactly that happen.
This will now probably forever be his angle in the popular media: Godfather of AI, mistaken, scoffed at by his peers. Not necessarily wrong, but not flattering either.
I can relate to the gut-wrenching feeling of seeing your ideas being bandied around as original, without any attribution. Of course someone in the field of reinforcement learning is obsessed with the credit assignment problem, no matter the time passed. How can AI researchers hope to solve this pressing problem, when they can't even assign credit to the originators of their ideas?
I wonder if this harsh piece would have been written the way it was, if the author realized that Schmidhuber may be on the spectrum, and that his erratic and obsessive behavior is an ailment, not a deliberate choice.
In the future, Schmidhuber will be mentioned in the same breath as Turing. Of his peers wishing he'd shut up, I don't have the same expectation.
From the headline, I knew it would be about Schmidhuber. Isn't this shtick getting a little old? I don't see how he is being suppressed, when I see him give talks and lead symposiums at conferences all the time ...
> The most prestigious AI conference goes by the unfortunate acronym of NIPS, or Neural Information Processing Systems.
This is the review process working as it should: helping situate work in a longer dialog so that readers can follow a chain of reasoning. Much of the rest of the drama discussed in this article is about allocating the celebrity that manifests when a technique suddenly becomes useful, through some combination of changes to the environment, theoretical improvements and reduction to practice. It's a very human concern, but it's more business (or politics) than science.
[1] ftp://ftp.idsia.ch/pub/juergen/factorial.pdf [2] http://papers.nips.cc/paper/5423-generative-adversarial-nets... [3] http://media.nips.cc/nipsbooks/nipspapers/paper_files/nips27...