English subtitles for clip: File:Will robots outsmart us.webm
Jump to navigation
Jump to search
1 00:00:00,160 --> 00:00:04,414 Will robots ever turn against us and take over the world? 2 00:00:04,414 --> 00:00:07,359 I'm going to tell you all about it. 3 00:00:08,000 --> 00:00:10,379 Will robots outsmart us? 4 00:00:16,139 --> 00:00:18,944 This the university of the Netherlands 5 00:00:18,944 --> 00:00:23,078 New technologies emerge quickly and movies show somehow... 6 00:00:23,078 --> 00:00:27,131 frightening scenarios in which robots are smarter than humans... 7 00:00:27,131 --> 00:00:28,954 and are taking over control. 8 00:00:28,954 --> 00:00:32,514 Some movies even go a step further and predict... 9 00:00:32,514 --> 00:00:34,480 that robots can replicate themselves. 10 00:00:35,120 --> 00:00:40,442 However will this ever become reality? Will robots outsmart us? 11 00:00:40,442 --> 00:00:44,299 As an applied mathematician and mechanical engineer... 12 00:00:44,299 --> 00:00:46,620 I'm trying to answer this question. 13 00:00:47,367 --> 00:00:52,061 Together with the team we work on recreating intelligence in robots. 14 00:00:52,061 --> 00:00:57,756 We strive to build robots that can learn, make decisions, recognize environments... 15 00:00:57,756 --> 00:01:01,231 and possibly create without human intervention. 16 00:01:01,231 --> 00:01:05,920 This is known as artificial intelligence or AI. 17 00:01:05,920 --> 00:01:11,938 If Ai can think as a human or basically if AI is conscious... 18 00:01:11,938 --> 00:01:15,114 then we call such AI general AI. 19 00:01:15,727 --> 00:01:21,280 On the other hand, if AI can replicate itself, then we call it super AI. 20 00:01:22,400 --> 00:01:26,020 However our society is far from achieving these two. 21 00:01:26,447 --> 00:01:31,202 Currently we mostly work on the so-called weak artificial intelligence. 22 00:01:31,202 --> 00:01:36,111 In weak AI we strive to use human reasoning... 23 00:01:36,111 --> 00:01:40,272 as the model for the robot's training, but not for its end goal. 24 00:01:40,272 --> 00:01:47,491 Basically we do not try to replicate full human mental capabilities in a robot... 25 00:01:47,491 --> 00:01:51,985 but instead we train robots to perform simple operations. 26 00:01:51,985 --> 00:01:56,314 For example that a robot is saying hi when somebody is passing by 27 00:01:56,314 --> 00:01:58,800 or that is laughing at on one of your jokes. 28 00:02:00,480 --> 00:02:04,989 At this moment probably most of you would think that AI and robotics... 29 00:02:04,989 --> 00:02:05,840 have the same goal. 30 00:02:06,640 --> 00:02:12,749 However this is not true. In robotics we strive to mimic physical human actions. 31 00:02:12,749 --> 00:02:16,964 We build robots that can pick up objects for example a ball... 32 00:02:16,964 --> 00:02:21,840 and place it on a specific position that is predefined by a human. 33 00:02:22,960 --> 00:02:25,999 In artificial intelligence on the other hand 34 00:02:25,999 --> 00:02:29,334 we are trying to recreate intelligence in robots. 35 00:02:29,334 --> 00:02:34,839 Basically we train robots to recognize objects, to understand that the object... 36 00:02:34,839 --> 00:02:40,203 is a ball and also to place them accordingly to a specific position... 37 00:02:40,203 --> 00:02:42,080 without human interaction. 38 00:02:43,840 --> 00:02:47,967 Therefore if we have AI and robotics together... 39 00:02:47,967 --> 00:02:50,720 then we get an artificially intelligent robot. 40 00:02:52,000 --> 00:02:57,208 In which AI acts as a brain and robotics acts as a body. 41 00:02:58,035 --> 00:03:01,269 So now you know what is artificial intelligence. 42 00:03:01,269 --> 00:03:06,447 However how can we train robots to be artificially intelligent? 43 00:03:06,447 --> 00:03:11,202 How can we train robots to make decisions, to recognize environments... 44 00:03:11,202 --> 00:03:16,638 or to act? Well for this we need to understand human behaviour... 45 00:03:16,638 --> 00:03:21,760 and therefore we need to understand how the human brain is functioning. 46 00:03:23,120 --> 00:03:27,411 by zooming in on the human brain, we see that it consists... 47 00:03:27,411 --> 00:03:29,520 of billions of brain cells. 48 00:03:30,720 --> 00:03:32,554 They are also called neurons. 49 00:03:32,554 --> 00:03:37,411 Each of these cells is an electrochemical structure. 50 00:03:37,411 --> 00:03:42,202 Basically each neuron starts with a set of input antennas... 51 00:03:42,202 --> 00:03:47,628 which are called dendrites. And in these antennas we receive signals... 52 00:03:47,628 --> 00:03:50,630 from the outside world or from the other cells. 53 00:03:50,631 --> 00:03:55,520 Once when the input signal is high enough... 54 00:03:56,320 --> 00:04:02,626 then the cell starts being activated and it's releasing a set of neurotransmitters... 55 00:04:02,626 --> 00:04:07,813 that travel right to the back of the cell to the synaps... 56 00:04:07,813 --> 00:04:11,040 which is a terminal for communication with another cell. 57 00:04:12,400 --> 00:04:17,208 On that moment we say that our cell is firing or being activated. 58 00:04:17,208 --> 00:04:24,140 Connecting cells into large networks, we can fire specific patterns... 59 00:04:24,140 --> 00:04:29,931 of brain cells and those are then understood as specific situations. 60 00:04:29,932 --> 00:04:35,200 For example if we have a spider in front of us, 61 00:04:35,840 --> 00:04:41,195 we will suddenly get afraid and some of us will run, the others will start screaming. 62 00:04:41,196 --> 00:04:48,999 Simply explained, this visual stimuli or spider is activating a set of cells... 63 00:04:48,999 --> 00:04:54,974 in our brain that are recognized as a pattern that means dangerous situation. 64 00:04:54,974 --> 00:05:00,852 In order to mimic this in robots, we have to understand this pattern... 65 00:05:00,852 --> 00:05:02,878 and we have to program it. 66 00:05:04,159 --> 00:05:10,240 To do so we actually mimic brain cells by artificial cells 67 00:05:10,800 --> 00:05:13,611 And then we connect them into large networks... 68 00:05:13,611 --> 00:05:15,520 which are similar to natural ones. 69 00:05:16,720 --> 00:05:22,515 However note that new artificial networks are not electrochemical devices. 70 00:05:22,515 --> 00:05:28,659 They are mathematical devices. Each cell is a mathematical element... 71 00:05:28,659 --> 00:05:32,080 with a set of inputs and a set of output signals 72 00:05:32,960 --> 00:05:35,522 If the input signal is strong enough 73 00:05:35,522 --> 00:05:39,692 then our mathematical element will get activated. 74 00:05:39,692 --> 00:05:44,761 It will release an output signal which then can activate some other cell. 75 00:05:46,174 --> 00:05:50,210 In this way we build patterns in large neural networks. 76 00:05:50,211 --> 00:05:55,056 However note that we need to train these patterns... 77 00:05:55,056 --> 00:05:57,347 because we need to train the neural network... 78 00:05:57,347 --> 00:05:59,871 that specific situations mean something. 79 00:05:59,871 --> 00:06:04,772 For example that a spider means dangerous situation. 80 00:06:04,960 --> 00:06:10,366 To do so we have to provide neural networks with huge amounts of data. 81 00:06:10,367 --> 00:06:16,497 For example in order that a robot learns that a spider means dangerous situation... 82 00:06:16,497 --> 00:06:21,455 we have to provide the robot with a lot of images of different spiders. 83 00:06:21,455 --> 00:06:27,120 In this way the robot can recognize features of an animal called spider. 84 00:06:29,040 --> 00:06:32,495 However in this way we are only training robots... 85 00:06:32,495 --> 00:06:36,080 to recognize objects, situations or environments. 86 00:06:37,200 --> 00:06:42,392 But not to act to it. For example if we see a spider we immediately run. 87 00:06:42,392 --> 00:06:45,195 So we would like that our robot does the same. 88 00:06:45,195 --> 00:06:52,137 To do so we have to actually understand human actions... 89 00:06:52,137 --> 00:06:59,760 in order to train robots to act and for this we are searching solutions in psychology. 90 00:07:01,280 --> 00:07:06,799 In 19th century Ivan Pavlov has studied something called classical conditioning. 91 00:07:07,680 --> 00:07:11,861 In his famous experiment he let the dog hear a bell sound... 92 00:07:11,861 --> 00:07:13,280 whenever he got some food. 93 00:07:14,240 --> 00:07:17,218 In this way the dog started drooling. 94 00:07:17,218 --> 00:07:22,331 Over time the association between food and bell sound became so strong... 95 00:07:22,331 --> 00:07:26,480 that actually the dog started drooling by only hearing the bell sound. 96 00:07:28,000 --> 00:07:30,848 And this is exactly conditioned learning. 97 00:07:30,848 --> 00:07:36,913 In conditioned learning we are learning an association between two stimuli... 98 00:07:36,913 --> 00:07:38,875 natural and neutral one. 99 00:07:38,875 --> 00:07:42,280 natural is food, neutral one is bell sound. 100 00:07:42,280 --> 00:07:48,871 Over time they are actually having a strong association among each other... 101 00:07:48,871 --> 00:07:51,120 and they will give the same response. 102 00:07:52,000 --> 00:07:56,213 Classical conditioning can also be translated to the robot's world. 103 00:07:56,213 --> 00:08:00,565 For example if we have a robot and two objects, car and a ball... 104 00:08:00,565 --> 00:08:03,402 and we would like that the robot is tracing a ball... 105 00:08:03,402 --> 00:08:08,758 then whenever the robot sees a ball, we actually can play a specific melody... 106 00:08:08,758 --> 00:08:12,640 that is implemented in the computer program as a reward. 107 00:08:13,520 --> 00:08:16,727 In this way whenever a robot is tracking a ball... 108 00:08:16,727 --> 00:08:19,119 he will know that he's doing good job. 109 00:08:21,440 --> 00:08:26,240 Over time the association between ball and melody will become so strong 110 00:08:26,240 --> 00:08:29,484 that actually the robot will start tracking any ball... 111 00:08:29,484 --> 00:08:31,919 without hearing any melody. 112 00:08:34,000 --> 00:08:38,001 In classical conditioning therefore we are learning association... 113 00:08:38,001 --> 00:08:44,308 between two stimuli: ball and melody, food and a bell sound. 114 00:08:44,308 --> 00:08:49,437 In machine learning this is known as prediction algorithm... 115 00:08:49,437 --> 00:08:52,918 in which we predict the outcome of each situation. 116 00:08:52,918 --> 00:08:57,520 However there is also a main issue in this learning technique 117 00:08:58,880 --> 00:09:02,989 If we repeat Pavlov's experiment and then give to a dog food... 118 00:09:02,989 --> 00:09:08,367 without any bell sound and if we do this repeatedly, the dog will forget... 119 00:09:08,367 --> 00:09:10,560 to drool whenever hearing a bell sound. 120 00:09:11,680 --> 00:09:14,479 Therefore this learning process can be diminished. 121 00:09:15,120 --> 00:09:18,616 And this is something we do not want to use for robots... 122 00:09:18,616 --> 00:09:22,319 because they will forget what they were trained for. 123 00:09:24,080 --> 00:09:27,223 Hence we need a more complex learning technique. 124 00:09:27,223 --> 00:09:30,079 And this is reinforcement learning. 125 00:09:30,960 --> 00:09:36,061 In reinforcement learning an animal or human will exhibit more frequently... 126 00:09:36,061 --> 00:09:40,805 an action that has led to a reward in the past and less frequently... 127 00:09:40,805 --> 00:09:42,800 actions that have led to punishments. 128 00:09:43,680 --> 00:09:48,829 For example, if you let a cat into a cage with four different buttons... 129 00:09:48,830 --> 00:09:54,480 and let's say that one of these buttons is providing a reward or food to a cat... 130 00:09:54,480 --> 00:09:58,453 and the three other buttons are representing punishments. 131 00:09:58,453 --> 00:10:04,841 For example by electrical current. Then after some time the cat will realize... 132 00:10:04,841 --> 00:10:09,589 That actually the fourth button is the one that provides food... 133 00:10:09,589 --> 00:10:13,269 and it will start using this button more often. 134 00:10:13,269 --> 00:10:17,535 This can be translated to the robot's world as well. 135 00:10:17,535 --> 00:10:23,188 In such a case we have to build an algorithm that is based on law of effect. 136 00:10:23,188 --> 00:10:28,610 That means that the algorithm has to compare different actions... 137 00:10:28,610 --> 00:10:32,817 by their outcomes and also has to build associations. 138 00:10:32,817 --> 00:10:37,280 It has to associate each situation with one pre-learned action. 139 00:10:39,120 --> 00:10:44,412 And this is allowing us then to train robots for unknown environments. 140 00:10:44,412 --> 00:10:47,465 For example if you want to send the robot to mars... 141 00:10:47,465 --> 00:10:52,080 then you would like that this robot can move without hitting any obstacles. 142 00:10:52,960 --> 00:10:55,730 And for this we use reinforcement learning. 143 00:10:55,730 --> 00:10:59,840 However the biggest issue in reinforcement learning 144 00:11:00,400 --> 00:11:03,429 is that it has to interact with an environment. 145 00:11:03,429 --> 00:11:06,319 We need to get feedback from the environment. 146 00:11:07,520 --> 00:11:09,697 And this is not always possible. 147 00:11:09,697 --> 00:11:13,744 For example we have environments that give us very sparse rewards... 148 00:11:13,744 --> 00:11:15,935 or no reward at all. 149 00:11:16,522 --> 00:11:21,680 For example if you want to train a robot to automatically play a computer game, 150 00:11:22,240 --> 00:11:26,603 then you can't use reinforcement learning if you are winning or losing... 151 00:11:26,603 --> 00:11:27,839 only in the end of the game. 152 00:11:30,160 --> 00:11:33,376 In such situations we need a new type of learning. 153 00:11:33,376 --> 00:11:35,680 This is called imitation learning 154 00:11:37,120 --> 00:11:42,661 As its name says in imitation learning we bring an imitator or teacher... 155 00:11:42,661 --> 00:11:46,511 that is actually demonstrating an action to a robot. 156 00:11:46,511 --> 00:11:50,201 In this way the robots gets more data... 157 00:11:50,201 --> 00:11:52,400 and the learning process becomes shorter. 158 00:11:53,680 --> 00:11:58,800 Imitation learning can be used for example if we send our robots to Mars. 159 00:11:58,800 --> 00:12:01,040 And we would like them to build settlements. 160 00:12:02,880 --> 00:12:05,262 By imitating human behaviour... 161 00:12:05,262 --> 00:12:09,839 robots will start building these structures in acollaborative manner. 162 00:12:10,880 --> 00:12:16,667 So now we are learning that actually it's not enough to train one robot individually... 163 00:12:16,667 --> 00:12:19,155 but we have to train them in a group. 164 00:12:19,155 --> 00:12:22,680 This is similar to human society. 165 00:12:22,681 --> 00:12:28,581 Humans are grouping together in order to collaborate for better survival. 166 00:12:28,581 --> 00:12:32,080 Therefore if we want to build human free factories... 167 00:12:32,640 --> 00:12:35,443 we have to use collaborative robotics. 168 00:12:35,443 --> 00:12:39,821 Imagine two robots that are welding together or two robots... 169 00:12:39,821 --> 00:12:44,650 that are printing material together into some very precise structure. 170 00:12:44,650 --> 00:12:47,477 In such case we can get a lot of issues. 171 00:12:47,477 --> 00:12:52,764 For example one robot could recognize an object and the other not. 172 00:12:52,764 --> 00:12:56,900 One is processing information much faster than the other. 173 00:12:56,900 --> 00:13:01,011 Then they are not aware of its own physical constraints... 174 00:13:01,011 --> 00:13:03,471 or physical constraints of other robots. 175 00:13:03,471 --> 00:13:06,743 And also they need to be aware of all actions.... 176 00:13:06,743 --> 00:13:09,629 that all other robots in the group are taking. 177 00:13:09,629 --> 00:13:14,320 Therefore we need to enhance AI to be collaborative 178 00:13:15,840 --> 00:13:19,840 It doesn't matter if we train an individual robot or a group of them. 179 00:13:20,400 --> 00:13:24,173 We still need to use big data sets to train robots. 180 00:13:24,173 --> 00:13:27,040 And therefore we need huge computing power. 181 00:13:27,920 --> 00:13:30,627 For this purpose we use supercomputers. 182 00:13:30,627 --> 00:13:34,159 However even a bee's brain is more powerful... 183 00:13:34,159 --> 00:13:36,938 than the largest supercomputer on earth. 184 00:13:36,938 --> 00:13:41,723 Of course supercomputers are faster in processing information. 185 00:13:41,723 --> 00:13:47,183 In this case this would be around 200 000 times faster than a bee's brain. 186 00:13:47,183 --> 00:13:50,960 However they require more energy. 187 00:13:51,920 --> 00:13:57,323 For processing the same information the bee's brain needs 10 microwatts of energy 188 00:13:57,323 --> 00:14:01,266 while the supercomputer needs 13 megawatts. 189 00:14:01,267 --> 00:14:06,804 Therefore AI algorithms require a lot of energy... 190 00:14:06,804 --> 00:14:12,420 and we need to build new ones that would be based only on small data sets... 191 00:14:12,420 --> 00:14:15,078 and hence use less energy. 192 00:14:15,078 --> 00:14:18,978 On the other hand when we train robots... 193 00:14:18,978 --> 00:14:22,387 we train them based on our previous knowledge. 194 00:14:22,388 --> 00:14:26,853 That means that a robot can make a new piece of music or a new drawing. 195 00:14:26,853 --> 00:14:30,319 However this is all based on our previous knowledge. 196 00:14:32,320 --> 00:14:35,156 And therefore they can't create anything new. 197 00:14:35,156 --> 00:14:41,771 Some researchers argue that then artificial intelligence is not intelligent at all. 198 00:14:41,772 --> 00:14:47,752 However I do believe that we will make robots more intelligent than they are today. 199 00:14:47,752 --> 00:14:51,824 For this we need to pass from building associative relationships... 200 00:14:51,824 --> 00:14:54,854 between stimuli or stimulus and action... 201 00:14:54,855 --> 00:14:58,124 and start building cause-effect relationships. 202 00:14:58,124 --> 00:15:02,425 For example, our robots today can very precisely move... 203 00:15:02,425 --> 00:15:06,400 but they don't know enough to predict possible accidents. 204 00:15:08,080 --> 00:15:12,189 Their knowledge is lower than knowledge of a 16 month old baby. 205 00:15:12,189 --> 00:15:18,645 Therefore we need to build AI that can build such cause-effect relationships. 206 00:15:18,645 --> 00:15:23,852 Today we have learned that the machine and robot intelligence... 207 00:15:23,852 --> 00:15:30,276 depends on big data sets, complex algorithms and huge computing power. 208 00:15:30,276 --> 00:15:35,430 However is this enough to achieve general or super AI? 209 00:15:35,431 --> 00:15:39,840 Well, only time will tell. However one thing is for sure. 210 00:15:40,720 --> 00:15:44,408 Robot intelligence will be different from human intelligence... 211 00:15:44,408 --> 00:15:49,065 in the same way as animal intelligence is different from human one. 212 00:15:49,065 --> 00:15:54,283 Robots will just excel in different types of skills than humans. 213 00:15:54,283 --> 00:15:57,890 And I do believe that one day they will make more... 214 00:15:57,890 --> 00:16:07,840 than what human species have done so far. Thank you.