Deep Learning

[Robert Goren]

I just read an article about deep learning in Mother Jones. Of course the left leaning writer was at level of paranoia that he could passed a conservative. But that is another story.
The question I have to the people here who keep up on this kind of stuff is "How close are they to make it work ?" and " How close is it becoming something that I or someone else working at home with or without the aid of cloud computing could employ?" and "Is it something that could effect the betting of horse racing at some point?"

[DJofSD]

RG, are there some links? I don't think I can fully understand your questions unless I have some background in which to frame them.

[Robert Goren]

RG, are there some links? I don't think I can fully understand your questions unless I have some background in which to frame them.Mother Jones has not posted it for non subscribers yet. They will eventually. It is titled "Do Androids Dream of Electric Lolcats?"
They are talking about the Google Brain project that taught itself to spot a cat from YouTube images. Andrew Ng was head of the project at the time. The usual suspects, Microsoft, Yahoo, Facebook, Netflix and of course the NSA are working "deep learning" projects. The article doesn't go into detail much. The writer, Dana Liebelson, seemed more interested in pointing the potential abuses. I figured if Mother Jones is reporting it, it must be old news in the tech community.

[DJofSD]

OK, thanks. If you can, whenever MJ does make it available, please update the thread.

I did a quick read on Wikipedia. A lot of what was there was about neural networks. So, to partially answer your questions, yes, deep learning can be applied to handicapping, and, there have been NN used in different software offerings for decades.

I've looked at NN as a self assignment for programming an app but that was long ago and I never finished it. In part, it was not finished because I kept on reading that one of the faults found with NNs is a regression to the mean. Whether or not that is correct, I'll leave to those with expertise.

[GameTheory]

NN are becoming popular again, and deep learning is mostly just a phrase for "NNs that don't suck like they used to". It seems to be much about doing more abstracting in layers so the problem can be broken down better, but really it is just a continuation of tech that has been around for decades and as computing power increases some old limitations are easier to get around.

The question "how close is it to working?" is a bit of a false one -- what would entail "getting it to work"?

"Strong AI" as a goal is fine but it is incredibly vague and improvements in these areas are incremental as they usually are, there is nothing revolutionary here. So it works as well as it works, it is coming along, improvements are made...but what's the threshold?

Can you use it? Sure, but programming required.

Check out http://numenta.com/ for what I think is more interesting work...