TED Talks: The wonderful and terrifying implications of computers that can learn - General

This video summarizes the current state of technology relating to machine learning. The speaker, Jeremy Howard, explains that most of the current mathematical foundation of machine learning (AKA Deep Learning) was developed by a single man years ago, and yet is still flexible enough to cover several different areas of "learning" including visual, audial, and abstract ideas. [video=youtube;t4kyRyKyOpo]http://www.youtube.com/watch?v=t4kyRyKyOpo[/video] As a programing hobbyist I found this to be quite interesting and something that I'd like to explore and practice on my own, however I figured some of you may find some of the practical examples in this video to be interesting. Although a debate of the ramifications of this technology would probably be best for a different thread

Neural networks and general Machine learning is fascinating, but this talk doesn't actually explain how they work at all which is the most interesting bit. Also saying 'computers have just learned to do this so they'll take jobs' is just hyperbole. Most of this stuff needs heavy computing power, and isn't as accurate. What we will probably see is better backups, ie autopilots that can adapt to weather conditions etc

I'm doing my Master's in Computational Biology so machine learning is very relevant. I wish things were less "mystified" though, because the image I get from the public is that machine learning & AI = GONNA KILL HUMANS. Machine learning isn't magic, neither is AI for that matter. The computer uses an algorithm to match patterns with an established database of patterns and then uses statistics to figure out the most likely match. That's more or less it. If the database doesn't exist, then we feed the algorithm a lot of data so that it can establish relationships between the data and assign probabilities to them. It's basically applied statistics.

I'm fine with AI. But we need more research about the impact.

[QUOTE=Occlusion;46737938]Neural networks and general Machine learning is fascinating, but this talk doesn't actually explain how they work at all which is the most interesting bit. Also saying 'computers have just learned to do this so they'll take jobs' is just hyperbole. Most of this stuff needs heavy computing power, and isn't as accurate. What we will probably see is better backups, ie autopilots that can adapt to weather conditions etc[/QUOTE] Computers are able to be more accurate than humans. We're not perfect. All we have is a neural network in our head, and that's precisely the foundation of this type of Machine Learning. The difference is: computers can process more things faster than you can. They can work faster than you can once they've learned, and they could continue to learn and become more accurate as they work (in much the same way that an experienced worker does). And a computer doesn't need to sleep or eat, and it certainly doesn't guzzle as much power as you seem to be imagining. In fact, compared to their human counterparts, they're pretty much the perfect candidate for a job application. That makes us squishy biological humans redundant. Read this book: [URL]http://www.amazon.co.uk/Superintelligence-Dangers-Strategies-Nick-Bostrom/dp/0199678111/ref=sr_1_1?ie=UTF8&qid=1418862425&sr=8-1&keywords=superintelligence[/URL] [IMG]https://dl.dropboxusercontent.com/u/10518681/Screenshots/2014-12-18_00-27-18.png[/IMG] Yes, it is all predictions. No, no one knows exactly when this will happen. Good Old-Fashioned AI (GOFAI), like the one used to play checkers in the 60's, has been around for ages. These deep learning techniques aren't much newer. But they are an indication of things to come. These demonstrations are something entertaining to gawk at, but what happens when machines can train and improve themselves? What would stop them becoming as intelligent as you or I? At what point do they no longer need us? And at that point, what happens to our jobs in a world which favours unstoppable, unremitting machines? [editline]18th December 2014[/editline] [QUOTE=Swebonny;46738011]I'm doing my Master's in Computational Biology so machine learning is very relevant. I wish things were less "mystified" though, because the image I get from the public is that machine learning & AI = GONNA KILL HUMANS. Machine learning isn't magic, neither is AI for that matter. The computer uses an algorithm to match patterns with an established database of patterns and then uses statistics to figure out the most likely match. That's more or less it. If the database doesn't exist, then we feed the algorithm a lot of data so that it can establish relationships between the data and assign probabilities to them. It's basically applied statistics.[/QUOTE] It's not magic, and that's why it's a real concern. Maybe things will turn out okay, but maybe you'd be worried if something came out tomorrow which had learned all he material you and I have on our University courses and could replace us? Give that book a read, it goes into technical details, you'd love it.

I feel like people need to know the difference between AI that can learn, and AI that can self-improve. AI that can learn has a limit, and not a particularly high one, it only learns what it's told to learn, basically. Self-Improving AI is what you're thinking of if you get afraid of learning AI.

[QUOTE=Occlusion;46737938]Also saying 'computers have just learned to do this so they'll take jobs' is just hyperbole. Most of this stuff needs heavy computing power, and isn't as accurate. What we will probably see is better backups, ie autopilots that can adapt to weather conditions etc[/QUOTE] not really i just think it's a matter of time. we've already seen rudimentary computers near enough replacing jobs in retail and warehouses and transport and production already like the self-driving cars that have been tested are already better than humans. all it needs is the infrastructure and pricing to make it happen. are they 100% perfect? no i guess not and they probably never will be. but they are damn better than humans, who manage to make cars one of the top ten most deadly things on the planet

If anything relevant to the societal implications of AIs wish to be covered, I'd go with CGP Grey's video (Humans Need Not Apply). Needless to say, informing the public on the actual science behind things is what prevents scaremongering and fantastical predictions like KILL ALL HUMANS.EXE and actually makes things interesting.

[QUOTE=LoneWolf_Recon;46740344]If anything relevant to the societal implications of AIs wish to be covered, I'd go with CGP Grey's video (Humans Need Not Apply). Needless to say, informing the public on the actual science behind things is what prevents scaremongering and fantastical predictions like KILL ALL HUMANS.EXE and actually makes things interesting.[/QUOTE] The science behind it gets pretty involved pretty quickly. That's why it's a difficult one. And plus, we're not at the level where self-improving ML algorithms exist - so what science should we be explaining to the public? All the ones in the video have been trained by a human, we're worried about ones that don't require human input to train. (When I say train, I mean training in the literal sense of training a neural network - maybe superintelligent machines wouldn't use neural networks, no one knows because we're not there yet) The stuff in the video is based around this: [url]http://en.wikipedia.org/wiki/Artificial_neural_network[/url]

I don't get why either the video or my post is getting rated dumb. What am I missing here? The speakers won't go into a too deep of a level of discussion because they have a limited time to speak. I just thought that the examples of current work utilizing the algorithm was interesting. I know it's not incredibly sophisticated, but it's a different level of computational programing than I had been exposed to before, I think.

[QUOTE=Swebonny;46738011]I'm doing my Master's in Computational Biology so machine learning is very relevant. I wish things were less "mystified" though, because the image I get from the public is that machine learning & AI = GONNA KILL HUMANS. Machine learning isn't magic, neither is AI for that matter. The computer uses an algorithm to match patterns with an established database of patterns and then uses statistics to figure out the most likely match. That's more or less it. If the database doesn't exist, then we feed the algorithm a lot of data so that it can establish relationships between the data and assign probabilities to them. It's basically applied statistics.[/QUOTE] It sounds so simple when you put it like that. I think its fascinating because our minds seem to function in almost the exact same way.