Moravec's Paradox - Why are machines so smart, yet so dumb?

this episode was made possible by Robo race the world's first driverless racing series hi everyone jto have you ever wondered why it's so easy for machines to do things we find really hard hey Siri what's the square root of five thousand six hundred and ninety seven this square root of five thousand six hundred and ninety seven is approximately seventy five point four seven eight five but so hard to do things we find easy a theory knock-knock no no no you say who's there Nana in the 80s a researcher named hands Larvik wondered the exact same thing putting it it is comparatively easy to make computers exhibit adult level performance on intelligence tests or playing checkers and difficult or impossible to give them the skills of a one-year-old when it comes to perception and mobility this curious observation became known as Mara vex paradox and even though it was formulated more than thirty years ago it's still very true today don't get me wrong AI has come a long way from beating well champions at the board game go the quiz game jeopardy the card game poker and the video game League of Legends but despite all of this it still has a hard time understanding a joke or navigating its way around a room without bumping into something today we're going to be exploring this paradox by looking at the two most mainstream approaches to creating AI hopefully by the end of the video you'll have a pretty good idea of what makes machines so smart and yet so dumb before we start a quick definition of artificial intelligence for those of you not so familiar with the phrase there are a lot of definitions out there but let's keep it simple artificial intelligence is the ability of a digital computer or computer controlled robot to perform tasks commonly associated with intelligent beings examples include problem-solving reacting to the environment working toward a goal and understanding language now let's begin by going all the way back to ancient Greece to a man who comes up a lot in the history of science Aristotle people had been reasoning logically for millennia but our subtle was the first to really analyze the process he identified a system where if two true facts were stated a third true fact followed logically for example all cats are adorable mia is a cat therefore mia is adorable what's important here is the form this was the first time anyone came up with a systematic way to automate a logical statement once you put in two related facts you only need to turn the crank and put out pops a third centuries later the self-taught mathematician and hardcore Aristotle fan George Boole expanded on Aristotle's ideas and invented what he called the laws of thought his idea was to translate all logical principles of human thought into mathematical statements for example his proposition for States the principle of contradiction affirms that it is impossible for any being to possess equality and at the same time to not possess it he translated this rule into the equation X multiplied by 1 minus x equals 0 where X represents a property like being furry 1 minus X represents its opposite and 0 represents a state of non-existence so this equation is saying a being that is both furry and not very does not exist bulls work showed that some kinds of logical reasoning could be carried out by simply manipulating equations these ideas were part of the roller-coaster ride to the belief that the nature of intelligence could be boiled down to nothing more than symbol manipulation that all human thought was simply manipulating symbols in our mind and if we could automate this process we could create a machine with the intelligence of a human this theory would become the foundation to one of the main approaches for creating an artificial intelligence because based on symbol manipulation it was simply called symbolic AI with the coming of the digital computer a machine that could perform said symbol manipulations the time was right to start seriously working toward an artificial intelligence in August 1955 a scientist named John McCarthy took the lead in organizing what was called a summer research project on artificial intelligence his proposal read we propose that a two-month ten month study of artificial intelligence be carried out during the summer of 1956 a Dartmouth College in Hanover New Hampshire the study is to proceed on the basis that every aspect of learning or any other feature of intelligence can be so precisely described that a machine can be made to simulate it there was an explosion of successes in the years that followed computer programs were solving algebra word problems proving theorems in geometry and learning to play checkers the money poured in and optimism rose lead AI researchers were proclaiming within a generation the problem of creating artificial intelligence will substantially be solved and machines will be capable within twenty years of doing anything a man can do this may sound crazy to us now as more than 60 years later machines are still very far from reaching human level intelligence so why were they so optimistic well if we look at all the problems that they were solving they're all hardwired in math and logic as these are difficult for people they assumed having solved for their hard problems that the easy ones like common sense and vision would just fall into place it's understandable why they thought this way if I asked you to name intelligent people from history he'd probably say names like Einstein da Vinci or Beethoven people who did things that the average person finds hard you probably wouldn't say your baby brother who figured out how to grab his rattle or your little niece who diggle's with surprise when you appeared out of thin air but as we've discovered since these are important aspects of intelligence and not once that symbolic AI with strong foundations in logic and reasoning are well equipped for to see why this is let's take an activity that most toddlers find easy building a tower of blocks you call your program builder you start with the command build but this command is much too complicated and must be broken down into simpler tasks it's like if in a chess game you simply commanded winning game so build is broken down into begin add and end but even add is too big a task and must be broken down first add must find a new block and then a hand must put the block on the tower all of these commands must be broken down again same must recognize what a block is no matter their size color place or if they're partially obscured move has to guide the hand through complicated parts in space yet never toppled the existing tower and how should we prevent grasp from grasping a block already in the tower how could find to determine what blocks are still available for use how would builder know if there are enough blocks to build a tall enough Tower what if the tower starts to sway builder would need to have a basic knowledge of physics to understand why when we look closely we find a bewildering world of questions that don't have straightforward answers compare this to the world of chess with its neat rules and clear goal we put the rules in the machine doesn't calculations and we get the best possible move out it's the same idea as with Aristotle's method of finding a third fact from two known facts sure we're putting a lot more in and the turning of the crank is a very complicated set of instructions but we provide all the information and we're at the set of instructions now don't get me wrong we've gotten pretty far doing this deep blue bead well chess champion garry kasparov doing this but we can see that its limitations lie in tasks that don't have clear well-defined rules that we can just pile in tasks in the real messy world if we imagine intelligence as a beautiful flower symbolic AI is if we just saw what was above the surface you know the pretty part and went straight to trying to create that it would resemble intelligence on the outside but it can't grow on its own it can't learn everything has to be added by hand it's not flexible and every task has to be treated individually in other words it doesn't contain any of the stuff that intelligence is actually made of now let's turn to another approach that has been seeing a lot of success in recent years neural networks neural networks take a completely different direction instead of replicating the visible pops of intelligence neural networks attempt to recreate the seed the stuff intelligence is actually made of and let it grow on its own as humans are the best example we have of intelligence it seemed logical to model them after the human brain in the last decade neural networks have revolutionized computer vision speech recognition medical image analysis natural language understanding and maybe the path towards self-driving cars that actually work but despite all of this we are still very far from a machine that resembles anything close to human level intelligence like c-3po from Star Wars or Samantha from the movie her that can carry a conversation and seem to be able to think for themselves to see why this is let's take a look at how neural networks work a neural network consists of well a network of neurons not real neurons of course but computational neurons each neuron has a weight associated with each connection the network is taught to do a task by having it analyzed training examples which have been previously labeled in advance a common example of a task for a neural network is an object recognition task where the neural network is presented with a large number of objects of a certain type such as pieces of cheese and by analyzing the recurring patterns in the images it learns to categorize new images by adjusting the weights on the connections you can get networks that do anything you like unlike symbolic AI neural networks aren't programmed directly for a task rather they have their requirements just like a child's developing brain then they learn the information a real-life example of where neural networks have allowed for amazing progress is in Robo races autonomous race cars who were kind enough to sponsor this video Robo race is the world's first ái driverless electric racing series a competition for human and machine teams using both self-driving and manually controlled cars Robo race is at its core a competition between human coders developers can download their driving software put it on a USB stick plug it into the car and the car will go for a drive by driving in simulations the engineers create thousands of scenarios and this data is fed to the car which can then learn to drive around a track however this is very different from how a human learns to drive as chief AI scientists Yan Laocoon says a human can learn to drive a car in 15 hours of training without crashing into anything if you want to use the current machine learning methods to train a car to drive itself the machine will have to drive off cliffs 10,000 times before it figures out how not to do that this is because a neural network needs thousands of samples of data to learn how to do a task well for a Rover car to drive on a track the track needs to be previously mapped out by a person and then the car must go through thousands of simulations before it's ready if you'd like to know more about Robo race and how their autonomous race cars work make sure to check out their show me how it works YouTube playlist at the end of this video which you can find in the description so if neural networks are modeled after the human brain why is there this difference well one reason is that even though they're inspired by the brain there are very very crude representation our brains contain about 86 billion neurons and more than 100 trillion connections whereas the number of neurons in an artificial network is much less than that one of the biggest ones being 60 million neurons large but more importantly there's still so much that we don't know about the human brain like how it learns or recalls information neural networks can now beat humans in pretty much any area where there's enough training data which can be turned into numerical values and the solutions to the examples are clear and well-defined but and this is a big but they can't generalize those skills a robot car could crash you in a race around a track but it most likely couldn't play a good game of poker plus a human needs to provide all the information to the neural network possibly hundreds of millions of data points no one's giving children labelled data points in that way most of the learning they do is before you can even tell them this is a piece of cheese joshua Tenenbaum a professor of computational cognitive science at MIT thinks the best approach to achieving human level intelligence is to focus on the start up software president infants things like intuitive physics and folk psychology he believes that the earlier an ingredient is present the more likely it is to be foundational to later learning and development so what do you think will machines ever be as intelligence as humans let me know your thoughts in the comments below hey Siri will you ever be as intelligent as a human I'm sorry I'm afraid I don't have an answer to that thank you to Robo race for sponsoring this episode of up and Adam if you'd like to support the channel and find out more about how their autonomous race cars work make sure to check out their playlist show me how it works which you can find in the description of this video and in the card subscribe to their Channel and add them on Instagram I hope you enjoyed the video and until next time bye hey Siri why did the chicken cross the road hey theory why did the chicken cross the road

Channel: Up and Atom

Views: 200,087

Keywords: morave's paradox, moravec, moravecs paradox, paradox, artificial intelligence, machine learning, supervised learning, neural network, unsupervised learning, computer science, convolutional neural network, deep learning, image classification, transfer learning, machine learning tutorial, artificial neural network, what is machine learning, data science, educational, computer vision, simplilearn machine learning, stochastic gradient descent, machine learning algorithms, roborace

Id: hcfVRkC3Dp0

Length: 14min 29sec (869 seconds)

Published: Mon Jul 08 2019