Teaching humans will be beneficial to only a few people that you can reach, but if you instead invest your time in teaching an ai, you can implicitly teach the whole of humanity…
The way to do this would be to come up with the algorithms that would then be learnt for deep learning etc.
so your brain has more than 10 billion neurons in it and the way it works is at each moment each year and has to decide whether to go ping on the basis that decision on pings it gets from other neurons and it waits those things so some things it takes a lot of notice of and these things tell me either you should go ping or you shouldn’t go ping and it changes those weights so by changing how much it listens to other neurons and your can change how it behaves and that’s how you learn everything
so that just leaves one question which is what’s the principle for changing how much you listen to other neurons and that’s called a learning algorithm and deep learning is a learning algorithm for changing how much one your arm will rely on other neurons to decide whether to coping device erm there’s a shallow learning as well oh yes the shallow loving that’s what the other people do and I’m that doesn’t have lots of layers of yards between the input and output so we’re in too deep learning here how does deep learning mimic how humans learn about the world well nobody really knows how in the real brain you change the strength of the connections that determine how much when you are in effect southern Europe but in the nineteen eighties people came up with a very effective algorithm for doing that and it’s meant to be a simplified model of the brain nobody knows if the brain actually works like this and back in the eighties people were very suspicious because the algorithm didn’t work that well but as computers got faster and we got bigger data sets this algorithm that works really well it’s used all over the place is used in your cell phone and so now it seems like a better bet for the brain might be up to you know who made it up this algorithm it was invented first in about nineteen seventy by some obscure guy it was reinvented by lots of people and then in the eighties when computers were faster after implemented effectively people started using it and showing what it could do but computers weren’t fast enough to make it really impressive then so mainstream a I didn’t believe in this algorithm what happened a few years ago with computers became fast enough and suddenly this algorithm started solving all the problems of mainstream a I couldn’t solve like recognizing speech for example would would Watson the computer from jeopardy we’ll beat everybody would that be part of what we’re talking about here there’s little bits of machine learning in watts and some of these bits may well use this algorithm but mostly it’s hand programming it’s a very impressive system that involves a huge amount of human labor to make it work and the idea of these artificial neural networks is you’re trying to learn everything
how does the artificial intelligence in watson compared to deep learning so the main difference is in deep learning you’re trying to learn everything with nobody program yet the only thing that gets programmed in your computer simulation is the learning algorithm huh everything inside this your own that gets learned from data not programmed in my home so they’re thinking um yes you could say that i just did yeah is that accurate though it’s independent thought in some respects you might irritate some philosophers but yes I think they are thinking
Neural network is a simulation of a whole bunch of neurons and it’s something that learns by changing the connections between neurons and is that what’s happening here so for recognizing the characters he uses in your match and that your match is trained on lots and lots of characters from lots of different fonts and with lots of different distortions and noise and a neural net is currently the best system for being able to write reliably recognize characters that are deformed and noisy now did this program just translate that because somebody made a code to consider every possible word in Spanish to translate or is this thing thinking okay for this particular program I think currently it’s not using your next to do the translation is using your nets to do the character recognition okay um but google and other people already have your nets doing translation and they’re doing translation they’re not being used online present and when you do google translate it all look at phrases in one language and translate them to phrase in the other language and this huge table but there’s a new way of doing machine translation that’s much more interesting that uses your nets where it reads the sentence in one language and turns it into a thought that is when i say something that expresses a thought and obviously the way to translation is to figure out the thought being expressed in the first language and say the same thing in the second language and google now has translation systems that work like that they’re about comparable with the existing translation system on a medium sized training set and they’re not quite as good as the existing system on a really big data set yet but they will be and in a few years time we’ll be doing machine translation by take the centers in one language turn it into a big pattern of your activity that is the thought behind that sentence and then save that thought in the other language can it understand nuance when it sees it it understands some news a present it it’s can use a lot of improvement still armed so the some things it can’t do a present like if i say to you in English the trophy would not fit in the suitcase because it was too big you know the itch refers to the trophy because it wouldn’t fit yes if I say the trophy would not fit in the suitcase because it was too small you know the it refers to the suitcase and that’s real world knowledge affecting how you translate now if you translate from English to French in French you just can’t say it you have to choose agenda yes and so we can’t translate that English sentence into the right Frank sent instead you need real world knowledge to decide what gender to make that it that will happen I don’t know if it will happen in a few years or 10 years but once that happens we’ll know that is really understanding and it can figure out homonyms without any difficulty stuff like that’s no problem that’s it is the use of complicated real world knowledge to disambiguate things huh and it’s beginning to be able to do a bit of contour appropriate is there one area in particular that you think deep learning is going to change the future um no I think it’s going to change the future and lots and lots of areas let me give you a few examples yes over the last few years it’s sort of become the method of choice for recognizing speech um it’s now becoming the method of choice for transcribing speech so going all the way from the sound wave to the transcript sure what said with just one your network that does everything it’s but going to become the method of choice for machine translation suppose you want to design a new drug you’d like to know um I i give you a bunch of Canada molecules and you like to know how well they’re buying to some target site and you like to predict that rather than doing the experiment is much cheaper to do a prediction the experiment and then you can experiment on the ones that are predicted to work well and neural Nets recently became the best method of doing that if you want to identify a pedestrian in the road and your let’s definitely the best method of doing that so it’s all over these these your nets especially the ones using this deep learning algorithms are going to be used everyone how many years away do you think we are from a neural network being able to do anything that a brain can do I don’t know it’s very hard to predict the future beyond previous I don’t think it will happen in the next five years beyond that it’s all kind of fog so I be very cautious about making a prediction is there anything about this that makes you nervous um in the very long run yes I mean obviously having other super intelligent beings or more intelligent than us is something to be nervous about it’s not going to happen for a long time but it is something to be nervous about him long run what aspect of it makes you nervous what will they be nice to us it’s just like the movies you’re worried about that scenario in the movies with a very long term yet when I think the next five or 10 years we don’t worry about it also the movies always portrayed as an individual intelligence I think it may be that it goes in a different direction where we sort of developed jointly with these things so the things aren’t fully autonomous that developed to help us there like personal assistants and will develop with them and it will be more of a symbiosis than our rivalry but we don’t set an expectation or a hope that’s not how it sounds like more hope than expectation let me read this you this is a from a piece in The Daily Beast last December by giclee Whitaker talking about the year a i took the wheel artificial intelligence did more than look at algorithms this year and while we’ve heard about super computers and quantum computing for years this is the first time that any of that lightning fast thinking out an answer texts started sharing the roads the roofs and the responsibilities with you and me and people are split over whether that was a good thing i do want to pursue this you know this notion of expectation vs hope you hope it will all work out well but in the long run I I sense a your expectation may not be quite as benign is that fair to say i think it’s very very hard to know what will happen beyond a five-year horizon so I my my state of mind is I just don’t know what’s going to happen I think trying to stop the technology will be very hard I mean if you look at automatic teller machines my guess is back when they were introduced people complained about them putting bank on the side of work but I think nobody i would say they were a bad idea even bank tellers even bank tell them i mean their job more interesting because they do with the tricky cases rather than you just want to say twenty dollars that right um so it’s clear that that technology is a force for good weather or technology is a force for good or for bad depends a lot on the political system or not the political system decides to do with it that’s what i wanted to follow up on because clearly things things in so many different areas of life are changing so quickly faster than our political systems are designed to make rules and laws around them so how deeply involved you think politics has to be our governments have to be in order to deal with the changes that are coming in this sector they’re going to have to be involved so if you just take driverless cars it’s pretty clear to everybody in the industry i think the driverless cars will save a whole lot of lives but the politicians are terrified of the first time driving this car run somebody down so politically if it driverless cars kill a few people to save tens of thousands of people that’s a problem for the politicians but they should just face up to it and say look these things are going to make as much safer it will take a brave politician to say I know two people were killed but here’s the 10,000 we saved you can’t see the 10,000 save you can certainly see the two killed yes so I know it’s gonna be a lot of that but it is very clear the driverless cars are going to be a good thing ok so in conclusion what kind of impact you hope deep learning has on our future I hope that it for example allows Google to read documents and understand what they say and so return much better search results to you so you can search by the content of the document rather by the words in the document I hope it’ll make for intelligent personal systems who arm can answer questions in a sensible way and have a sensible conversation as opposed to a conversation that keeps getting derailed it’ll give us driverless cars that’s clearly gonna come pretty soon it’ll make computers much easier to use I think because you’ll be able to just say to your computer how do i print this damn thing and the computer will do it rather you have to figure out all these commands so it should make if you’re right it should make our lives better yes it should be just like automatic teller machines that we should make that little bit of life better but it should do them for a lot of things fingers crossed