Can machines Think Like Humans? | Samuel Bosch | TEDxEcublens

what do you think of when you hear terms like machine learning or artificial intelligence well I happen to know just the right person to help us hey Siri what is machine learning machine learning is the scientific study of algorithms and statistical models that computer systems use to perform as ok but actually that was a very precise answer to my question just way too technical for this talk highly intelligent machine learning algorithms are getting most of the media coverage these days for example self-driving cars or virtual assistants just like Siri over here this evening though I'm going to show you a completely different type of machine learning instead of focusing on the most intelligent algorithms I'm going to show you the power of keeping things small and simple did you ever try calling Siri when your phone wasn't connected to the Internet well here's what happens hey Siri what is machine learning Siri doesn't work without internet connection and the reason it won't work is because our our phones not have their our computing power required for advanced natural language processing in other words our phones are simply way too slow for performing such calculations all of our hey Siri requests I actually process in a huge data center online by Apple if you try to run Siri offline on your phone it would be extremely slow and it would drain your batteries probably within minutes at the same time our human brains seem to be way ahead of supercomputers at such tasks we don't struggle answering simple questions and we can even deal with much more complicated ones the big question now is can we make computers as energy efficient as human brains even better what if we could create algorithms which work in the same way the human brain does well it turns out we can do it at least in some applications but it does not paste our neural net as some of you might speculate I started working on a project to address this question last year at University of California in San Diego in the group of Professor Tyler Rosen our goal was to create an algorithm which works in the same way in a similar way the human brain does so we wanted to create an algorithm which could perform machine learning offline on small scale sometimes even battery-powered devices for example cell phones smartwatches or even TVs blenders vacuum cleaners you name it can anyone without using a calculator tell me what the first five decimals are of square root of 17 I don't think anybody here can do it but a simple calculator can do this within split seconds how is this possible if our brains are supposedly so much more advanced than the most powerful supercomputer this is because our brains are not evolutionary made to calculate the numbers numbers are not natural to our brains hundreds of thousands of years ago our ancestors survival didn't depend on their ability to solve complex complicated numerical problems with high accuracy their survival depended on their ability to hunt to make quick intuitive decisions based on what they observed and heard and also the ability to do advanced math wasn't exactly helpful in attracting females guess it hasn't really changed but you know is this a smart is the new sexy so I guess I can still pass therefore we need to create machine learning algorithms which don't rely on mathematics in the same way other algorithms do so how do we do this we do it by recognizing patterns together with my colleagues at UC San Diego we managed to create an incredibly energy efficient algorithm which can perform machine learning offline on small scale battery powered devices we call it hyper dimensional computing or HD computing the basic idea is this to give the computer a bunch of training all in Asia for example pictures of cats and dogs from the internet which you can see on the top and then the algorithm learns like a little child so it takes these pictures these images encodes them using some random numbers and converts them into patterns so we have one pattern representing dogs as one pattern representing cats very simple now when we give the computer a new unknown image like the one in the bottom it is also converted into a pattern and then the algorithm simply checks if it is closer to the pattern of a dog or a cat and in this particular case it is going to tell us that it's a super cute little puppy and of course it is it's my family's dog milky you might be asking yourself now what is this all good for why can't we just connect to the internet and run algorithms the same ways Theory does to answer this question I'm going to show you a project created by the Berkeley Wireless Research Center here you can see a man using hand gestures to control a prosthetic arm has a small chip sorry a small detector attached to his right arm just over here you can see it to the right of me and this this chip measures electric impulses from his nervous system now those impulses are then transmitted to a chip which uses machine learning HD computing to be precise to predict which next move his arm is going to make that information is then passed along to the prosthetic arm which you can see on the right which says then gets into this position now here's the trick if that man didn't have the bottom half of his arm if the bottom half of his arm was missing the detector would still be able to tell which which move his arm would make if it was there and in the case of a disabled person that prosthetic arm is then attached to his body that way he or she can use a prosthetic arm as if it was a real and almost fully functional arm sounds amazing right so let us watch this in action pay close attention to the fingers of the man into the fingers of the prosthetic arm this is of course just an early prototype developed at Berkeley but it is still a remarkable proof of concepts we want such prosthetic limbs to work reliably and without requiring concept internet connection why well just imagine a man with a prosthetic leg crossing a large street and suddenly stopping in the middle of it then a police officer comes over and starts yelling at him to get off the street because of all the incoming traffic the man's response would be well I'm really sorry officer but my leg just lost internet connection maybe you could share me a mobile hotspot for a second obviously such things don't work without efficient offline processing and also we don't want live data of our bodies in the hands of large international corporations right I would like you to leave today remembering that artificial intelligence and machine learning are not just about making computers more intelligent in humans there's so much potential in equipping small-scale battery-powered devices with machine learning algorithms as Professor Yan rabbi from Berkeley said artificial intelligence is one of the most profound steps in evolution of humankind AI will help us adapt to the rapidly changing environment around us in some way it will help humanity cope with a smart world that is emerging as we speak in conclusion by embracing artificial intelligence we empower positive change for society thank you very much [Applause]

Channel: TEDx Talks

Views: 24,345

Rating: 4.9432626 out of 5

Keywords: TEDxTalks, English, Technology, Computers, Data Science, Engineering, Future, Internet, Research

Id: IP0HXomV4j8

Length: 8min 37sec (517 seconds)

Published: Fri Dec 13 2019