1 00:00:01,680 --> 00:00:03,669 covering the week's top tech stories 2 00:00:03,669 --> 00:00:03,679 covering the week's top tech stories 3 00:00:03,679 --> 00:00:07,190 covering the week's top tech stories with a slight linux bias a deep learning 4 00:00:07,190 --> 00:00:07,200 with a slight linux bias a deep learning 5 00:00:07,200 --> 00:00:08,870 with a slight linux bias a deep learning model has achieved superhuman 6 00:00:08,870 --> 00:00:08,880 model has achieved superhuman 7 00:00:08,880 --> 00:00:11,509 model has achieved superhuman performance at gran turismo sport 8 00:00:11,509 --> 00:00:11,519 performance at gran turismo sport 9 00:00:11,519 --> 00:00:13,910 performance at gran turismo sport over the past few decades research teams 10 00:00:13,910 --> 00:00:13,920 over the past few decades research teams 11 00:00:13,920 --> 00:00:15,509 over the past few decades research teams worldwide have developed machine 12 00:00:15,509 --> 00:00:15,519 worldwide have developed machine 13 00:00:15,519 --> 00:00:16,230 worldwide have developed machine learning 14 00:00:16,230 --> 00:00:16,240 learning 15 00:00:16,240 --> 00:00:17,990 learning and deep learning techniques that can 16 00:00:17,990 --> 00:00:18,000 and deep learning techniques that can 17 00:00:18,000 --> 00:00:20,070 and deep learning techniques that can achieve human comparable performance on 18 00:00:20,070 --> 00:00:20,080 achieve human comparable performance on 19 00:00:20,080 --> 00:00:21,990 achieve human comparable performance on a variety of tasks 20 00:00:21,990 --> 00:00:22,000 a variety of tasks 21 00:00:22,000 --> 00:00:23,830 a variety of tasks in order to further assess their 22 00:00:23,830 --> 00:00:23,840 in order to further assess their 23 00:00:23,840 --> 00:00:25,990 in order to further assess their capabilities and performances 24 00:00:25,990 --> 00:00:26,000 capabilities and performances 25 00:00:26,000 --> 00:00:27,750 capabilities and performances some of these models were also trained 26 00:00:27,750 --> 00:00:27,760 some of these models were also trained 27 00:00:27,760 --> 00:00:29,669 some of these models were also trained to play renowned board or video games 28 00:00:29,669 --> 00:00:29,679 to play renowned board or video games 29 00:00:29,679 --> 00:00:29,990 to play renowned board or video games such 30 00:00:29,990 --> 00:00:30,000 such 31 00:00:30,000 --> 00:00:33,030 such as the ancient chinese game go or atari 32 00:00:33,030 --> 00:00:33,040 as the ancient chinese game go or atari 33 00:00:33,040 --> 00:00:35,990 as the ancient chinese game go or atari arcade games researchers at university 34 00:00:35,990 --> 00:00:36,000 arcade games researchers at university 35 00:00:36,000 --> 00:00:38,389 arcade games researchers at university of zurich and sony ai zurich have 36 00:00:38,389 --> 00:00:38,399 of zurich and sony ai zurich have 37 00:00:38,399 --> 00:00:40,069 of zurich and sony ai zurich have recently tested the performance of a 38 00:00:40,069 --> 00:00:40,079 recently tested the performance of a 39 00:00:40,079 --> 00:00:41,670 recently tested the performance of a deep reinforcement 40 00:00:41,670 --> 00:00:41,680 deep reinforcement 41 00:00:41,680 --> 00:00:43,670 deep reinforcement learning-based approach that was trained 42 00:00:43,670 --> 00:00:43,680 learning-based approach that was trained 43 00:00:43,680 --> 00:00:45,750 learning-based approach that was trained to play the car racing video game gran 44 00:00:45,750 --> 00:00:45,760 to play the car racing video game gran 45 00:00:45,760 --> 00:00:47,350 to play the car racing video game gran turismo sport 46 00:00:47,350 --> 00:00:47,360 turismo sport 47 00:00:47,360 --> 00:00:49,190 turismo sport their findings further highlight the 48 00:00:49,190 --> 00:00:49,200 their findings further highlight the 49 00:00:49,200 --> 00:00:50,950 their findings further highlight the potential of deep learning techniques 50 00:00:50,950 --> 00:00:50,960 potential of deep learning techniques 51 00:00:50,960 --> 00:00:52,709 potential of deep learning techniques for controlling cars in simulated 52 00:00:52,709 --> 00:00:52,719 for controlling cars in simulated 53 00:00:52,719 --> 00:00:54,150 for controlling cars in simulated environments 54 00:00:54,150 --> 00:00:54,160 environments 55 00:00:54,160 --> 00:00:56,229 environments juan long song one of the researchers 56 00:00:56,229 --> 00:00:56,239 juan long song one of the researchers 57 00:00:56,239 --> 00:00:57,270 juan long song one of the researchers who carried out 58 00:00:57,270 --> 00:00:57,280 who carried out 59 00:00:57,280 --> 00:01:00,150 who carried out the studies says autonomous driving at 60 00:01:00,150 --> 00:01:00,160 the studies says autonomous driving at 61 00:01:00,160 --> 00:01:02,229 the studies says autonomous driving at high speed is a challenging task that 62 00:01:02,229 --> 00:01:02,239 high speed is a challenging task that 63 00:01:02,239 --> 00:01:04,549 high speed is a challenging task that requires generating fast and precise 64 00:01:04,549 --> 00:01:04,559 requires generating fast and precise 65 00:01:04,559 --> 00:01:05,509 requires generating fast and precise actions 66 00:01:05,509 --> 00:01:05,519 actions 67 00:01:05,519 --> 00:01:07,270 actions even when the vehicle is approaching its 68 00:01:07,270 --> 00:01:07,280 even when the vehicle is approaching its 69 00:01:07,280 --> 00:01:09,750 even when the vehicle is approaching its physical limits autonomous car racing 70 00:01:09,750 --> 00:01:09,760 physical limits autonomous car racing 71 00:01:09,760 --> 00:01:11,510 physical limits autonomous car racing where the goal is to complete a given 72 00:01:11,510 --> 00:01:11,520 where the goal is to complete a given 73 00:01:11,520 --> 00:01:13,429 where the goal is to complete a given course in minimal time 74 00:01:13,429 --> 00:01:13,439 course in minimal time 75 00:01:13,439 --> 00:01:15,030 course in minimal time features some of these difficulties of 76 00:01:15,030 --> 00:01:15,040 features some of these difficulties of 77 00:01:15,040 --> 00:01:16,870 features some of these difficulties of controlling a car close to its physical 78 00:01:16,870 --> 00:01:16,880 controlling a car close to its physical 79 00:01:16,880 --> 00:01:18,070 controlling a car close to its physical limitations 80 00:01:18,070 --> 00:01:18,080 limitations 81 00:01:18,080 --> 00:01:20,070 limitations to address these challenges and advance 82 00:01:20,070 --> 00:01:20,080 to address these challenges and advance 83 00:01:20,080 --> 00:01:21,990 to address these challenges and advance the frontier we consider the task of 84 00:01:21,990 --> 00:01:22,000 the frontier we consider the task of 85 00:01:22,000 --> 00:01:23,990 the frontier we consider the task of autonomous car racing in the top selling 86 00:01:23,990 --> 00:01:24,000 autonomous car racing in the top selling 87 00:01:24,000 --> 00:01:25,270 autonomous car racing in the top selling car racing game 88 00:01:25,270 --> 00:01:25,280 car racing game 89 00:01:25,280 --> 00:01:27,670 car racing game gran turismo sport which is known for 90 00:01:27,670 --> 00:01:27,680 gran turismo sport which is known for 91 00:01:27,680 --> 00:01:29,670 gran turismo sport which is known for its detailed physics simulation of 92 00:01:29,670 --> 00:01:29,680 its detailed physics simulation of 93 00:01:29,680 --> 00:01:32,950 its detailed physics simulation of various cars and tracks 94 00:01:32,950 --> 00:01:32,960 various cars and tracks 95 00:01:32,960 --> 00:01:35,270 various cars and tracks the key objective of the study is to 96 00:01:35,270 --> 00:01:35,280 the key objective of the study is to 97 00:01:35,280 --> 00:01:37,030 the key objective of the study is to develop an artificial neural network 98 00:01:37,030 --> 00:01:37,040 develop an artificial neural network 99 00:01:37,040 --> 00:01:38,230 develop an artificial neural network based controller 100 00:01:38,230 --> 00:01:38,240 based controller 101 00:01:38,240 --> 00:01:40,230 based controller that can autonomously navigate a race 102 00:01:40,230 --> 00:01:40,240 that can autonomously navigate a race 103 00:01:40,240 --> 00:01:41,910 that can autonomously navigate a race car within a simulated track 104 00:01:41,910 --> 00:01:41,920 car within a simulated track 105 00:01:41,920 --> 00:01:44,069 car within a simulated track without requiring any prior knowledge of 106 00:01:44,069 --> 00:01:44,079 without requiring any prior knowledge of 107 00:01:44,079 --> 00:01:45,670 without requiring any prior knowledge of the car's dynamics 108 00:01:45,670 --> 00:01:45,680 the car's dynamics 109 00:01:45,680 --> 00:01:47,670 the car's dynamics to perform well at gran turismo sport 110 00:01:47,670 --> 00:01:47,680 to perform well at gran turismo sport 111 00:01:47,680 --> 00:01:49,109 to perform well at gran turismo sport the controller should be trying to 112 00:01:49,109 --> 00:01:49,119 the controller should be trying to 113 00:01:49,119 --> 00:01:50,789 the controller should be trying to minimize the amount of time 114 00:01:50,789 --> 00:01:50,799 minimize the amount of time 115 00:01:50,799 --> 00:01:53,830 minimize the amount of time in which it can complete a given track 116 00:01:53,830 --> 00:01:53,840 in which it can complete a given track 117 00:01:53,840 --> 00:01:56,230 in which it can complete a given track the researchers first defined a reward 118 00:01:56,230 --> 00:01:56,240 the researchers first defined a reward 119 00:01:56,240 --> 00:01:58,069 the researchers first defined a reward function that formulated the racing 120 00:01:58,069 --> 00:01:58,079 function that formulated the racing 121 00:01:58,079 --> 00:01:58,789 function that formulated the racing problem 122 00:01:58,789 --> 00:01:58,799 problem 123 00:01:58,799 --> 00:02:01,270 problem as a minimum time problem as well as 124 00:02:01,270 --> 00:02:01,280 as a minimum time problem as well as 125 00:02:01,280 --> 00:02:03,109 as a minimum time problem as well as outlining a neural network policy that 126 00:02:03,109 --> 00:02:03,119 outlining a neural network policy that 127 00:02:03,119 --> 00:02:05,109 outlining a neural network policy that directly mapped input observations to 128 00:02:05,109 --> 00:02:05,119 directly mapped input observations to 129 00:02:05,119 --> 00:02:06,550 directly mapped input observations to car control commands 130 00:02:06,550 --> 00:02:06,560 car control commands 131 00:02:06,560 --> 00:02:08,309 car control commands subsequently they trained their neural 132 00:02:08,309 --> 00:02:08,319 subsequently they trained their neural 133 00:02:08,319 --> 00:02:09,669 subsequently they trained their neural network's parameters 134 00:02:09,669 --> 00:02:09,679 network's parameters 135 00:02:09,679 --> 00:02:12,070 network's parameters using reinforcement learning maximizing 136 00:02:12,070 --> 00:02:12,080 using reinforcement learning maximizing 137 00:02:12,080 --> 00:02:13,430 using reinforcement learning maximizing the reward that their model would 138 00:02:13,430 --> 00:02:13,440 the reward that their model would 139 00:02:13,440 --> 00:02:14,390 the reward that their model would receive when 140 00:02:14,390 --> 00:02:14,400 receive when 141 00:02:14,400 --> 00:02:16,949 receive when when performing well the researchers 142 00:02:16,949 --> 00:02:16,959 when performing well the researchers 143 00:02:16,959 --> 00:02:18,550 when performing well the researchers trained the neural network on four 144 00:02:18,550 --> 00:02:18,560 trained the neural network on four 145 00:02:18,560 --> 00:02:21,430 trained the neural network on four playstation 4s and a desktop pc 146 00:02:21,430 --> 00:02:21,440 playstation 4s and a desktop pc 147 00:02:21,440 --> 00:02:24,150 playstation 4s and a desktop pc remarkably after less than 73 hours of 148 00:02:24,150 --> 00:02:24,160 remarkably after less than 73 hours of 149 00:02:24,160 --> 00:02:24,869 remarkably after less than 73 hours of training 150 00:02:24,869 --> 00:02:24,879 training 151 00:02:24,879 --> 00:02:26,710 training their model had already achieved 152 00:02:26,710 --> 00:02:26,720 their model had already achieved 153 00:02:26,720 --> 00:02:28,390 their model had already achieved superhuman performance 154 00:02:28,390 --> 00:02:28,400 superhuman performance 155 00:02:28,400 --> 00:02:30,630 superhuman performance outperforming the best known human lap 156 00:02:30,630 --> 00:02:30,640 outperforming the best known human lap 157 00:02:30,640 --> 00:02:33,030 outperforming the best known human lap times in all three reference settings 158 00:02:33,030 --> 00:02:33,040 times in all three reference settings 159 00:02:33,040 --> 00:02:35,190 times in all three reference settings including two different cars on two 160 00:02:35,190 --> 00:02:35,200 including two different cars on two 161 00:02:35,200 --> 00:02:36,550 including two different cars on two different tracks 162 00:02:36,550 --> 00:02:36,560 different tracks 163 00:02:36,560 --> 00:02:38,550 different tracks the results highlight the incredible 164 00:02:38,550 --> 00:02:38,560 the results highlight the incredible 165 00:02:38,560 --> 00:02:40,390 the results highlight the incredible level of precision that the model can 166 00:02:40,390 --> 00:02:40,400 level of precision that the model can 167 00:02:40,400 --> 00:02:41,190 level of precision that the model can achieve 168 00:02:41,190 --> 00:02:41,200 achieve 169 00:02:41,200 --> 00:02:43,509 achieve and its ability to consistently execute 170 00:02:43,509 --> 00:02:43,519 and its ability to consistently execute 171 00:02:43,519 --> 00:02:45,110 and its ability to consistently execute optimal actions 172 00:02:45,110 --> 00:02:45,120 optimal actions 173 00:02:45,120 --> 00:02:46,949 optimal actions even in scenarios where these actions 174 00:02:46,949 --> 00:02:46,959 even in scenarios where these actions 175 00:02:46,959 --> 00:02:49,190 even in scenarios where these actions are risky and might lead a human player 176 00:02:49,190 --> 00:02:49,200 are risky and might lead a human player 177 00:02:49,200 --> 00:02:51,190 are risky and might lead a human player to push their car off track 178 00:02:51,190 --> 00:02:51,200 to push their car off track 179 00:02:51,200 --> 00:02:54,550 to push their car off track so what exactly would you reward an ai 180 00:02:54,550 --> 00:02:54,560 so what exactly would you reward an ai 181 00:02:54,560 --> 00:02:55,270 so what exactly would you reward an ai with 182 00:02:55,270 --> 00:02:55,280 with 183 00:02:55,280 --> 00:02:59,509 with chips what i really love about this 184 00:02:59,509 --> 00:02:59,519 chips what i really love about this 185 00:02:59,519 --> 00:03:01,270 chips what i really love about this story that becca's sharing with us 186 00:03:01,270 --> 00:03:01,280 story that becca's sharing with us 187 00:03:01,280 --> 00:03:04,229 story that becca's sharing with us is that when i think about ai technology 188 00:03:04,229 --> 00:03:04,239 is that when i think about ai technology 189 00:03:04,239 --> 00:03:06,070 is that when i think about ai technology up until this point when it comes to 190 00:03:06,070 --> 00:03:06,080 up until this point when it comes to 191 00:03:06,080 --> 00:03:07,670 up until this point when it comes to that machine learning 192 00:03:07,670 --> 00:03:07,680 that machine learning 193 00:03:07,680 --> 00:03:10,790 that machine learning it's very like step based 194 00:03:10,790 --> 00:03:10,800 it's very like step based 195 00:03:10,800 --> 00:03:13,910 it's very like step based so you think about chess for example so 196 00:03:13,910 --> 00:03:13,920 so you think about chess for example so 197 00:03:13,920 --> 00:03:15,670 so you think about chess for example so when an ai plays chess 198 00:03:15,670 --> 00:03:15,680 when an ai plays chess 199 00:03:15,680 --> 00:03:18,390 when an ai plays chess it's planning out the next move and then 200 00:03:18,390 --> 00:03:18,400 it's planning out the next move and then 201 00:03:18,400 --> 00:03:19,350 it's planning out the next move and then it makes that move 202 00:03:19,350 --> 00:03:19,360 it makes that move 203 00:03:19,360 --> 00:03:21,350 it makes that move and then it plans out the next move and 204 00:03:21,350 --> 00:03:21,360 and then it plans out the next move and 205 00:03:21,360 --> 00:03:22,470 and then it plans out the next move and then it makes that move 206 00:03:22,470 --> 00:03:22,480 then it makes that move 207 00:03:22,480 --> 00:03:24,149 then it makes that move what's happening here in gran turismo 208 00:03:24,149 --> 00:03:24,159 what's happening here in gran turismo 209 00:03:24,159 --> 00:03:25,589 what's happening here in gran turismo sport is that 210 00:03:25,589 --> 00:03:25,599 sport is that 211 00:03:25,599 --> 00:03:28,309 sport is that the ai the machine learning algorithm 212 00:03:28,309 --> 00:03:28,319 the ai the machine learning algorithm 213 00:03:28,319 --> 00:03:28,630 the ai the machine learning algorithm the 214 00:03:28,630 --> 00:03:28,640 the 215 00:03:28,640 --> 00:03:31,630 the the brains of the operation has to 216 00:03:31,630 --> 00:03:31,640 the brains of the operation has to 217 00:03:31,640 --> 00:03:33,270 the brains of the operation has to continually 218 00:03:33,270 --> 00:03:33,280 continually 219 00:03:33,280 --> 00:03:36,309 continually be planning ahead and figuring out 220 00:03:36,309 --> 00:03:36,319 be planning ahead and figuring out 221 00:03:36,319 --> 00:03:39,589 be planning ahead and figuring out how to make those moves and and it does 222 00:03:39,589 --> 00:03:39,599 how to make those moves and and it does 223 00:03:39,599 --> 00:03:41,110 how to make those moves and and it does it so very well 224 00:03:41,110 --> 00:03:41,120 it so very well 225 00:03:41,120 --> 00:03:44,309 it so very well so it makes me think okay first of all 226 00:03:44,309 --> 00:03:44,319 so it makes me think okay first of all 227 00:03:44,319 --> 00:03:47,110 so it makes me think okay first of all this could be a really really great way 228 00:03:47,110 --> 00:03:47,120 this could be a really really great way 229 00:03:47,120 --> 00:03:48,550 this could be a really really great way to improve 230 00:03:48,550 --> 00:03:48,560 to improve 231 00:03:48,560 --> 00:03:52,390 to improve the computer players in online games or 232 00:03:52,390 --> 00:03:52,400 the computer players in online games or 233 00:03:52,400 --> 00:03:55,670 the computer players in online games or in offline games imagine being able 234 00:03:55,670 --> 00:03:55,680 in offline games imagine being able 235 00:03:55,680 --> 00:03:58,390 in offline games imagine being able like to play a race car game with 236 00:03:58,390 --> 00:03:58,400 like to play a race car game with 237 00:03:58,400 --> 00:03:59,270 like to play a race car game with opponents that 238 00:03:59,270 --> 00:03:59,280 opponents that 239 00:03:59,280 --> 00:04:02,630 opponents that match your level of skill 240 00:04:02,630 --> 00:04:02,640 match your level of skill 241 00:04:02,640 --> 00:04:05,670 match your level of skill and maybe even surpass it 242 00:04:05,670 --> 00:04:05,680 and maybe even surpass it 243 00:04:05,680 --> 00:04:08,869 and maybe even surpass it but by using those computer games as 244 00:04:08,869 --> 00:04:08,879 but by using those computer games as 245 00:04:08,879 --> 00:04:12,149 but by using those computer games as simulations they're training the ai 246 00:04:12,149 --> 00:04:12,159 simulations they're training the ai 247 00:04:12,159 --> 00:04:15,270 simulations they're training the ai even further to learn what works in a 248 00:04:15,270 --> 00:04:15,280 even further to learn what works in a 249 00:04:15,280 --> 00:04:16,949 even further to learn what works in a simulated environment where 250 00:04:16,949 --> 00:04:16,959 simulated environment where 251 00:04:16,959 --> 00:04:20,949 simulated environment where human lives are never at risk so 252 00:04:20,949 --> 00:04:20,959 human lives are never at risk so 253 00:04:20,959 --> 00:04:23,830 human lives are never at risk so would they then be able to take that 254 00:04:23,830 --> 00:04:23,840 would they then be able to take that 255 00:04:23,840 --> 00:04:25,030 would they then be able to take that learned 256 00:04:25,030 --> 00:04:25,040 learned 257 00:04:25,040 --> 00:04:27,749 learned methodology that learned intelligence if 258 00:04:27,749 --> 00:04:27,759 methodology that learned intelligence if 259 00:04:27,759 --> 00:04:28,950 methodology that learned intelligence if you will 260 00:04:28,950 --> 00:04:28,960 you will 261 00:04:28,960 --> 00:04:31,670 you will and move it over to true autonomous 262 00:04:31,670 --> 00:04:31,680 and move it over to true autonomous 263 00:04:31,680 --> 00:04:33,670 and move it over to true autonomous vehicles that are going to operate 264 00:04:33,670 --> 00:04:33,680 vehicles that are going to operate 265 00:04:33,680 --> 00:04:37,749 vehicles that are going to operate on the road and are able to 266 00:04:37,749 --> 00:04:37,759 on the road and are able to 267 00:04:37,759 --> 00:04:41,189 on the road and are able to basically think for themselves and plan 268 00:04:41,189 --> 00:04:41,199 basically think for themselves and plan 269 00:04:41,199 --> 00:04:42,950 basically think for themselves and plan how are they going to pass a car without 270 00:04:42,950 --> 00:04:42,960 how are they going to pass a car without 271 00:04:42,960 --> 00:04:44,950 how are they going to pass a car without actually hitting that car 272 00:04:44,950 --> 00:04:44,960 actually hitting that car 273 00:04:44,960 --> 00:04:46,790 actually hitting that car and those kinds of things that right now 274 00:04:46,790 --> 00:04:46,800 and those kinds of things that right now 275 00:04:46,800 --> 00:04:48,790 and those kinds of things that right now we're just not quite there yet 276 00:04:48,790 --> 00:04:48,800 we're just not quite there yet 277 00:04:48,800 --> 00:04:50,629 we're just not quite there yet with autonomous vehicles but could this 278 00:04:50,629 --> 00:04:50,639 with autonomous vehicles but could this 279 00:04:50,639 --> 00:04:53,909 with autonomous vehicles but could this be kind of a precursor to the next step 280 00:04:53,909 --> 00:04:53,919 be kind of a precursor to the next step 281 00:04:53,919 --> 00:04:56,629 be kind of a precursor to the next step of artificial intelligence based 282 00:04:56,629 --> 00:04:56,639 of artificial intelligence based 283 00:04:56,639 --> 00:04:58,629 of artificial intelligence based autonomous vehicles and then that leads 284 00:04:58,629 --> 00:04:58,639 autonomous vehicles and then that leads 285 00:04:58,639 --> 00:04:58,950 autonomous vehicles and then that leads to 286 00:04:58,950 --> 00:04:58,960 to 287 00:04:58,960 --> 00:05:01,270 to everything else i think about space 288 00:05:01,270 --> 00:05:01,280 everything else i think about space 289 00:05:01,280 --> 00:05:02,150 everything else i think about space flight and and 290 00:05:02,150 --> 00:05:02,160 flight and and 291 00:05:02,160 --> 00:05:04,710 flight and and everything that is to come it's really 292 00:05:04,710 --> 00:05:04,720 everything that is to come it's really 293 00:05:04,720 --> 00:05:24,230 everything that is to come it's really an exciting time 294 00:05:24,230 --> 00:05:24,240 295 00:05:24,240 --> 00:05:26,320 you