fchollet

Doesn't sound like an economically rational move.

I would certainly be curious to know how much patio11 is really making from his consulting business. It sounds like he would be making at least $300,000 (a lowball estimate of $x0,000 would be $15,000, and a lowball estimate of how many weeks he works in a year would be 20). And possible up to $1.2M (40 weeks at $30k).

What an interesting world that would be, a world where we would still be burning witches. You wouldn't be typing this on a computer musing about ending aging; computers would not exist and neither would you.

Iterating endlessly on a single recipe is just not how you build a sustainable artistic business. It makes money in the near term, but it is doomed in the long term. The music labels and Hollywood will have to realize this at some point.

And because they're terrible at this type of optimization problem, they're getting stuck in a local minimum (the space is anything but convex), which happens to be repetitive, awful, and targeted at the largest common denominator.

Treating music as an optimization problem may not be a bad thing in itself, the issue is being unable to solve the problem properly.

Still, following it would require some familiarity with Numpy and scikit-learn (or other libs in the same spirit). As well as some experience with neural networks.

If you raise this issue on the mailing list or in the Github discussion, you'll get more help and advice.

When it comes to convolution, it will actually use cuDNN if available, and performance will be slightly better than Torch and Caffe: https://github.com/soumith/convnet-benchmarks#layer-wise-ben...

You want to incentivize the reporting of verified infractions, while de-incentivize false reports. A consensus system could be used to "verify" a report.

[1] http://www.fool.com/investing/general/2014/04/18/why-the-dow...

Each input is propagated through the unit (all timesteps) before the gradient is computed. No RTRL. It works well in practice.

Such science "journalism" is seriously hurting the field, including the researchers that are lionized in these articles. It also makes large groups of researchers who've been working hard on advancing the field think they don't matter in the eyes of the public, and that can be quite demoralizing.

That "unsupervised cats" paper from 2012 had nothing new in it at the time, although it was a feat of parallelization on a cluster of CPUs. Three years later its (uninteresting) results are not used anywhere. But Google used it as a PR piece at the time, and hundreds of journalists have been presenting it as some kind of game-changing breakthrough. It continues to this day.

If there is no evidence either way, yet every other developer raves about types all the time (the way "connaisseurs" rave about fine wine and expensive scotch despite being unable to objectively tell the difference), doesn't it follow that types are overrated?

A fairly easy way to introduce rotation invariance in DCNNS is to perform random rotations on the inputs during training. Likewise for scale invariance. Translation invariance is already introduced by the convolution operation itself.

The thing about deep learning, is that transform kernels are learned, not pre-computed as in classical signal processing. A DCNN will learn whatever convolution kernels it needs to perform the task at hand. I wouldn't be surprised if a DCNN trained on a classical signal processing task ended up rediscovering some well-known transform kernels originally derived from physical first principles...

timer.start();

computation();

time = timer.stop();

sleep(random_centered_on(mean-time));

Chances are it is already happening. More advanced AI will give governments and corporations the ability to do the same in a much more effective way, and on a larger scale. Large-scale intelligent data mining will make it possible to use people's data to build actionable models of what they think, what they will do next, and how to affect what they think and do. Better than humans could.

It doesn't even need to occur through sockpuppets, so the anti-sockpuppet regulation you propose would be not only highly intrusive but also ineffective. Here's an example: Facebook can manipulate your emotional state by selecting what goes into your newsfeed [1].

[1] http://www.theguardian.com/technology/2014/jun/29/facebook-u...

Meanwhile, we are completely missing the discussion we need to have about the realistic, short and medium-term dangers associated with the development of AI. Intelligent automation is about to disrupt economic production and existing power balances, much like software has done not too long ago, with significant social consequences. This is what we need to be talking about, and preparing for. We need to plan for a smooth transition to a post-AI world.

Regulating AI for fear that it may take over makes about as much sense as outlawing space travel for fear of aliens. Can we start having a sane discussion about AI now?

Yann LeCun: "Some people have asked what would prevent a hypothetical super-intelligent autonomous benevolent A.I. to “reprogram” itself and remove its built-in safeguards against getting rid of humans. Most of these people are not themselves A.I. researchers, or even computer scientists.[...] There is no truth to that perspective if we consider the current A.I. research. Most people do not realize how primitive the systems we build are, and unfortunately, many journalists (and some scientists) propagate a fear of A.I. which is completely out of proportion with reality. We would be baffled if we could build machines that would have the intelligence of a mouse in the near future, but we are far even from that." http://www.popsci.com/bill-gates-fears-ai-ai-researchers-kno...

Yoshua Bengio: "What people in my field do worry about is the fear-mongering that is happening [...] As researchers, we have an obligation to educate the public about the difference between Hollywood and reality."http://www.popsci.com/why-artificial-intelligence-will-not-o...

Rodney Brooks: "I say relax everybody. If we are spectacularly lucky we’ll have AI over the next thirty years with the intentionality of a lizard, and robots using that AI will be useful tools. And they probably won’t really be aware of us in any serious way. Worrying about AI that will be intentionally evil to us is pure fear mongering. And an immense waste of time."http://www.rethinkrobotics.com/artificial-intelligence-tool-...

Michael Littman: "Let's get one thing straight: A world in which humans are enslaved or destroyed by superintelligent machines of our own creation is purely science fiction. Like every other technology, AI has risks and benefits, but we cannot let fear dominate the conversation or guide AI research." http://www.livescience.com/49625-robots-will-not-conquer-hum...

To be fair, another major contributor to Torch is a co-author of this paper (Kavukcuoglu).

It's the fear-mongering that's the issue. It's as if these same pundits were warning us about the dangers of space travel because it could hypothetically cause us (1000 years from now?) to draw the attention of a dangerous alien civilization (does that even exist?) that could destroy the Earth. It's the same level of ridiculous speculation. And that has no place in the scientific discourse.

Write sci-fi novels if you care about this issue, but don't pretend it's science, much less a pressing technological issue.

"AI doom" is a powerful psychological trope for people in technology, and much like the hypothetical rogue super-AIs, it has gone out of control.

Andrej is a brilliant researcher, currently doing his PhD. He has a bright career ahead.

Maybe you should actually read his comment instead of dismissing it crudely, and likewise for the thoughts of the likes of LeCun, Ng, Bengio, etc. These are the people I would listen to, not the Nostradamus pundits.