0 00:00:00,000 --> 00:00:30,000 Dear viewer, these subtitles were generated by a machine via the service Trint and therefore are (very) buggy. If you are capable, please help us to create good quality subtitles: https://c3subtitles.de/talk/1393 Thanks! 1 00:00:19,680 --> 00:00:20,760 OK, so 2 00:00:22,230 --> 00:00:24,629 maybe you are always 3 00:00:24,630 --> 00:00:26,879 wondered how you could do Jedi mind 4 00:00:26,880 --> 00:00:29,189 tricks with a computer, and 5 00:00:29,190 --> 00:00:31,049 that's exactly why we're here now. 6 00:00:31,050 --> 00:00:33,329 So Agnew is going to 7 00:00:33,330 --> 00:00:35,639 tell you the fundamentals 8 00:00:35,640 --> 00:00:38,609 of EEG based brain computer interface, 9 00:00:38,610 --> 00:00:40,679 and he's always been 10 00:00:40,680 --> 00:00:43,529 fascinated with the human brain and 11 00:00:43,530 --> 00:00:45,629 he is a researcher and that scope 12 00:00:45,630 --> 00:00:47,820 and I give the stage to you. 13 00:00:50,280 --> 00:00:51,280 Hello. 14 00:00:53,850 --> 00:00:55,919 And the reason why I'm giving this talk 15 00:00:55,920 --> 00:00:58,169 is recently there 16 00:00:58,170 --> 00:01:00,179 has been a development and 17 00:01:00,180 --> 00:01:02,279 electroencephalography that was 18 00:01:02,280 --> 00:01:04,680 developed about 100 years ago 19 00:01:05,850 --> 00:01:08,159 and has been used in research 20 00:01:08,160 --> 00:01:09,640 and in medicine as well. 21 00:01:10,950 --> 00:01:12,030 But we now have 22 00:01:14,100 --> 00:01:16,469 consumer grade EEG headsets, 23 00:01:16,470 --> 00:01:18,689 as well as some open hardware 24 00:01:18,690 --> 00:01:20,909 projects aiming 25 00:01:20,910 --> 00:01:23,459 to develop EEG headsets. 26 00:01:25,200 --> 00:01:27,419 There I have a picture of the 27 00:01:27,420 --> 00:01:29,459 amount of EPOC, which I think was the 28 00:01:29,460 --> 00:01:31,769 first consumer great EEG 29 00:01:31,770 --> 00:01:33,299 headset. 30 00:01:33,300 --> 00:01:35,639 And actually, 31 00:01:35,640 --> 00:01:37,919 I think the aim of the Open 32 00:01:37,920 --> 00:01:40,679 BCI project is to 33 00:01:40,680 --> 00:01:42,749 get cheap 34 00:01:42,750 --> 00:01:44,370 research grade hardware. 35 00:01:47,420 --> 00:01:49,579 I'm not going to explain too much 36 00:01:49,580 --> 00:01:51,949 about about 37 00:01:51,950 --> 00:01:52,950 the devices. 38 00:01:54,680 --> 00:01:57,379 I want to talk a bit about 39 00:01:57,380 --> 00:01:59,599 how we can use EEG 40 00:01:59,600 --> 00:02:01,669 readings to have to have 41 00:02:01,670 --> 00:02:03,079 a brain computer interface. 42 00:02:06,600 --> 00:02:08,099 A brain computer interface 43 00:02:09,780 --> 00:02:12,059 typically consists of a user having 44 00:02:12,060 --> 00:02:14,340 a task, the task can be 45 00:02:16,650 --> 00:02:18,749 thinking, for example, to it 46 00:02:18,750 --> 00:02:20,069 to have some input to 47 00:02:21,120 --> 00:02:23,489 if if it's used to 48 00:02:23,490 --> 00:02:25,199 to drive an electric wheelchair, for 49 00:02:25,200 --> 00:02:26,999 example, it could be Atholl to go 50 00:02:27,000 --> 00:02:28,000 forward. 51 00:02:29,860 --> 00:02:32,099 The signal has to be the EEG signal 52 00:02:32,100 --> 00:02:33,659 has to be acquired. 53 00:02:33,660 --> 00:02:35,999 I'm not going to talk about 54 00:02:36,000 --> 00:02:38,129 that. I'm more focusing on the 55 00:02:38,130 --> 00:02:39,669 pre processing of the data and the 56 00:02:39,670 --> 00:02:42,239 feature. Extraction classification 57 00:02:42,240 --> 00:02:44,009 can generally be done with all kinds of 58 00:02:44,010 --> 00:02:47,009 classifiers popular 59 00:02:47,010 --> 00:02:48,889 support vector machines. 60 00:02:48,890 --> 00:02:50,789 Um, but a good feature. 61 00:02:50,790 --> 00:02:52,619 Extraction is essential. 62 00:02:53,850 --> 00:02:55,080 We cannot really do 63 00:02:57,000 --> 00:02:58,589 machine learning approaches where we 64 00:02:58,590 --> 00:03:01,049 learn the features because we typically 65 00:03:01,050 --> 00:03:02,999 do not have very much training data 66 00:03:04,860 --> 00:03:06,629 doing e.g. experiments with human 67 00:03:06,630 --> 00:03:09,029 subjects. Takes a lot of time 68 00:03:09,030 --> 00:03:10,199 and also 69 00:03:12,240 --> 00:03:13,949 the data might contain 70 00:03:15,450 --> 00:03:17,159 private information. 71 00:03:17,160 --> 00:03:19,919 So often, the datasets are not 72 00:03:19,920 --> 00:03:20,920 made public. 73 00:03:24,570 --> 00:03:26,639 So yeah, that's what I'm going to talk 74 00:03:26,640 --> 00:03:28,589 about. Mainly, but generally after 75 00:03:28,590 --> 00:03:29,929 classification, we have an open 76 00:03:29,930 --> 00:03:31,529 translation that can be a virtual 77 00:03:31,530 --> 00:03:33,719 keyboard or something, and 78 00:03:33,720 --> 00:03:37,109 a can be that is optional, a feedback 79 00:03:37,110 --> 00:03:39,329 that allows the user to a trained brain 80 00:03:39,330 --> 00:03:40,389 computer interface. 81 00:03:42,000 --> 00:03:44,069 Um, here 82 00:03:44,070 --> 00:03:46,139 I'm showing you the timeline 83 00:03:46,140 --> 00:03:47,829 of an EEG signal. 84 00:03:47,830 --> 00:03:51,149 This was a resting state experiment. 85 00:03:51,150 --> 00:03:53,189 So the subject was just resting. 86 00:03:53,190 --> 00:03:55,709 Doing nothing was eyes open. 87 00:03:55,710 --> 00:03:56,710 Um, 88 00:03:57,960 --> 00:03:59,969 you can't see very much on the table and 89 00:03:59,970 --> 00:04:03,179 it looks quite random. 90 00:04:03,180 --> 00:04:05,039 Random oscillations quite low. 91 00:04:05,040 --> 00:04:07,589 Generally, the signals are in some 92 00:04:07,590 --> 00:04:10,599 milli void range. 93 00:04:10,600 --> 00:04:12,689 Um, so one of the first 94 00:04:12,690 --> 00:04:14,819 steps that you can do to a time 95 00:04:14,820 --> 00:04:16,889 frequency analysis. 96 00:04:16,890 --> 00:04:19,349 Um, so what we have here is we have 14 97 00:04:19,350 --> 00:04:22,199 seconds of EEG 98 00:04:22,200 --> 00:04:24,459 and 14 electrodes 99 00:04:24,460 --> 00:04:26,789 that the 14 channels as the 100 00:04:26,790 --> 00:04:29,099 emotive epoch AEG's headset 101 00:04:29,100 --> 00:04:30,119 has. 102 00:04:30,120 --> 00:04:32,759 Um. And here we also have 14 seconds. 103 00:04:32,760 --> 00:04:35,009 And this is basically 104 00:04:35,010 --> 00:04:36,779 doing a 105 00:04:38,580 --> 00:04:40,799 couple of get 106 00:04:40,800 --> 00:04:41,800 computing 107 00:04:43,170 --> 00:04:45,299 spectrum for different time 108 00:04:45,300 --> 00:04:47,399 slots. So you see the development 109 00:04:47,400 --> 00:04:49,169 of the spectrum over time. 110 00:04:49,170 --> 00:04:50,969 And what you see here is one of the 111 00:04:50,970 --> 00:04:53,099 things that make it difficult. 112 00:04:53,100 --> 00:04:54,570 Most of the signal power 113 00:04:56,520 --> 00:04:58,859 is in the range below 114 00:04:59,970 --> 00:05:00,970 five hertz. 115 00:05:07,450 --> 00:05:09,039 The different frequency ranges are 116 00:05:09,040 --> 00:05:10,420 typically associated with, 117 00:05:11,950 --> 00:05:14,679 um, for example, 118 00:05:14,680 --> 00:05:17,259 some states of mind like sleep 119 00:05:17,260 --> 00:05:18,789 states. 120 00:05:18,790 --> 00:05:20,589 Actually, it's quite easy even with the 121 00:05:20,590 --> 00:05:22,869 timeline to to 122 00:05:22,870 --> 00:05:24,970 see and which sleep state someone is 123 00:05:29,500 --> 00:05:31,869 important. This is the 124 00:05:31,870 --> 00:05:32,870 alphabet, and 125 00:05:34,180 --> 00:05:37,359 that's something that should be in. 126 00:05:37,360 --> 00:05:39,189 And the plot that I showed before as 127 00:05:39,190 --> 00:05:41,469 well, um, 128 00:05:41,470 --> 00:05:42,699 we couldn't really see that in the 129 00:05:42,700 --> 00:05:44,889 timeline if actually I had instructed 130 00:05:44,890 --> 00:05:47,229 the subject to have the eyes closed 131 00:05:47,230 --> 00:05:48,969 but not open. 132 00:05:48,970 --> 00:05:51,549 Um, we would have seen oscillations 133 00:05:51,550 --> 00:05:54,129 in that range on the 134 00:05:54,130 --> 00:05:56,349 electrodes that are 135 00:05:56,350 --> 00:05:58,419 above the visual cortex 136 00:05:58,420 --> 00:06:00,609 because they would 137 00:06:00,610 --> 00:06:02,709 get all to idle state as if 138 00:06:02,710 --> 00:06:03,909 nothing is seen. 139 00:06:03,910 --> 00:06:06,039 And we would have more power 140 00:06:06,040 --> 00:06:07,040 in the 141 00:06:08,410 --> 00:06:10,539 UM if we are doing e.g. 142 00:06:10,540 --> 00:06:12,039 experiments. 143 00:06:12,040 --> 00:06:14,589 Um, we 144 00:06:14,590 --> 00:06:16,539 typically look into changes because we 145 00:06:16,540 --> 00:06:19,149 have this huge random noise signal 146 00:06:20,350 --> 00:06:22,449 where we have basically 147 00:06:22,450 --> 00:06:24,819 no idea what it means. 148 00:06:24,820 --> 00:06:26,889 We typically have experiments where we 149 00:06:26,890 --> 00:06:28,269 look at some, um, 150 00:06:29,660 --> 00:06:32,019 at at least two different states. 151 00:06:33,100 --> 00:06:35,319 So we define something as 152 00:06:35,320 --> 00:06:37,959 a baseline and then we typically 153 00:06:37,960 --> 00:06:40,059 look at what changes if we, for 154 00:06:40,060 --> 00:06:42,549 example, we have a resting state task 155 00:06:42,550 --> 00:06:43,749 for the subject. 156 00:06:43,750 --> 00:06:46,479 And then the task is to think 157 00:06:46,480 --> 00:06:49,329 maybe the um, 158 00:06:49,330 --> 00:06:51,819 the command to move the VHF. 159 00:06:51,820 --> 00:06:53,889 Um, what we see 160 00:06:53,890 --> 00:06:56,469 here is again, from the same 161 00:06:56,470 --> 00:06:58,059 EEG recording. 162 00:06:58,060 --> 00:06:59,060 Um, 163 00:07:00,130 --> 00:07:03,099 what I did was I used the 164 00:07:03,100 --> 00:07:05,679 two seconds before that I plotted here, 165 00:07:05,680 --> 00:07:07,989 but it looks about the same 166 00:07:07,990 --> 00:07:10,179 uh, computed the average and 167 00:07:11,440 --> 00:07:13,599 um, divided it all by 168 00:07:13,600 --> 00:07:14,619 it. 169 00:07:14,620 --> 00:07:16,809 We call that baseline 170 00:07:16,810 --> 00:07:17,829 correction. 171 00:07:17,830 --> 00:07:18,850 So now 172 00:07:20,590 --> 00:07:22,899 it's easier to see changes here and 173 00:07:22,900 --> 00:07:24,699 in the other areas. 174 00:07:24,700 --> 00:07:26,949 So having 175 00:07:26,950 --> 00:07:29,230 a baseline is something that you 176 00:07:30,670 --> 00:07:32,739 normally do if you have some EEG 177 00:07:32,740 --> 00:07:33,740 experiments. 178 00:07:37,360 --> 00:07:40,389 We also have some other problems besides 179 00:07:40,390 --> 00:07:42,579 General Becnel background noise 180 00:07:42,580 --> 00:07:43,580 that we have. 181 00:07:44,230 --> 00:07:46,009 It's artifacts here. 182 00:07:46,010 --> 00:07:47,010 It's a similar 183 00:07:48,970 --> 00:07:50,639 timeline. 184 00:07:50,640 --> 00:07:52,779 The difference is the 185 00:07:52,780 --> 00:07:54,879 subject was instructed to blink 186 00:07:54,880 --> 00:07:56,049 and intervals. 187 00:07:56,050 --> 00:07:57,799 You see that there are some huge peaks 188 00:07:57,800 --> 00:07:59,859 there, especially on 189 00:07:59,860 --> 00:08:01,539 the lower and the higher lines. 190 00:08:01,540 --> 00:08:03,729 That's because, according to 191 00:08:03,730 --> 00:08:05,800 the 10 20 system, 192 00:08:07,240 --> 00:08:09,429 the electrodes are numbered around the 193 00:08:09,430 --> 00:08:11,559 head. So on the top and on the 194 00:08:11,560 --> 00:08:13,929 bottom, it's basically the electrodes 195 00:08:13,930 --> 00:08:15,249 on the forehead. 196 00:08:15,250 --> 00:08:18,519 So what we see there is I artifacts 197 00:08:18,520 --> 00:08:20,619 actually to see 198 00:08:20,620 --> 00:08:21,620 artifacts. 199 00:08:22,910 --> 00:08:25,539 That's quite easy in the timeline. 200 00:08:25,540 --> 00:08:28,299 The problem is getting rid of it 201 00:08:28,300 --> 00:08:29,499 if we don't want to have it. 202 00:08:29,500 --> 00:08:31,929 So we start typically by instructing 203 00:08:31,930 --> 00:08:34,479 subjects not to blink, 204 00:08:34,480 --> 00:08:35,480 not to move. 205 00:08:37,059 --> 00:08:39,279 Also, a problem 206 00:08:39,280 --> 00:08:42,399 can be having. 207 00:08:42,400 --> 00:08:44,469 For example, the 50 hard power 208 00:08:44,470 --> 00:08:45,470 grid 209 00:08:47,570 --> 00:08:49,719 can also be an artifact in the scanner. 210 00:08:49,720 --> 00:08:51,070 Therefore, we typically have 211 00:08:54,370 --> 00:08:55,549 filters at 212 00:08:56,710 --> 00:08:57,710 its frequency. 213 00:09:00,160 --> 00:09:02,080 There are different approaches to 214 00:09:03,160 --> 00:09:04,269 get rid of the artifacts. 215 00:09:04,270 --> 00:09:06,309 The simplest one is cutting out of the 216 00:09:06,310 --> 00:09:07,310 data. 217 00:09:08,440 --> 00:09:10,599 The parts with artifacts. 218 00:09:10,600 --> 00:09:13,569 But this basically means we have to 219 00:09:13,570 --> 00:09:15,819 repeat the experiment to have it several 220 00:09:15,820 --> 00:09:16,820 times. 221 00:09:19,840 --> 00:09:21,279 But we are doing that anyway. 222 00:09:22,400 --> 00:09:24,040 Um, if we. 223 00:09:30,450 --> 00:09:32,999 One approach for brain computer interface 224 00:09:33,000 --> 00:09:35,220 are event related potentials. 225 00:09:36,450 --> 00:09:38,309 If we have an event that could be a 226 00:09:38,310 --> 00:09:41,279 subject is shown a picture 227 00:09:41,280 --> 00:09:42,809 or any other stimuli, 228 00:09:45,540 --> 00:09:46,829 and we repeat that. 229 00:09:48,360 --> 00:09:50,549 And then we everard's 230 00:09:50,550 --> 00:09:52,799 of all the reputations of showing this 231 00:09:52,800 --> 00:09:53,800 image. 232 00:09:55,050 --> 00:09:57,369 Then all the random noise will 233 00:09:58,710 --> 00:09:59,789 cancel out. 234 00:09:59,790 --> 00:10:02,489 And what we are left with is 235 00:10:02,490 --> 00:10:04,619 the EEG component that actually 236 00:10:04,620 --> 00:10:06,809 depends on the processing 237 00:10:06,810 --> 00:10:08,820 of showing this image 238 00:10:10,230 --> 00:10:12,419 and that we call an event relentlessly 239 00:10:12,420 --> 00:10:13,649 related potential. 240 00:10:14,820 --> 00:10:16,769 This is just an example. 241 00:10:16,770 --> 00:10:18,389 Typically, it doesn't look that nice. 242 00:10:21,850 --> 00:10:23,909 We typically columns the peaks, 243 00:10:23,910 --> 00:10:24,929 so we have three 244 00:10:26,070 --> 00:10:28,229 positive peaks here we call P1, 245 00:10:28,230 --> 00:10:30,869 P2, P3 and the P3 246 00:10:30,870 --> 00:10:33,479 also called P3 100 because it's 247 00:10:33,480 --> 00:10:35,459 about 300 milliseconds after the 248 00:10:35,460 --> 00:10:36,369 stimulus. 249 00:10:36,370 --> 00:10:37,979 That is something that we use for brain 250 00:10:37,980 --> 00:10:39,120 computer interfaces. 251 00:10:40,800 --> 00:10:41,800 Um, 252 00:10:43,290 --> 00:10:45,689 because there 253 00:10:45,690 --> 00:10:48,179 is the so-called oddball paradigm. 254 00:10:48,180 --> 00:10:50,789 The P3 100 is only there 255 00:10:50,790 --> 00:10:53,159 if something is relevant to your task 256 00:10:53,160 --> 00:10:55,229 and it's not happening very 257 00:10:55,230 --> 00:10:56,219 often. 258 00:10:56,220 --> 00:10:58,379 So the P300 speller, which 259 00:10:58,380 --> 00:11:00,659 is something like a virtual 260 00:11:00,660 --> 00:11:02,699 keyboard. Unfortunately, I don't have an 261 00:11:02,700 --> 00:11:03,700 animation. 262 00:11:04,650 --> 00:11:06,989 The lines and rows would light 263 00:11:06,990 --> 00:11:08,900 up at at 264 00:11:09,990 --> 00:11:11,159 some speed. 265 00:11:11,160 --> 00:11:13,229 If you want to type a letter, 266 00:11:13,230 --> 00:11:15,449 you would focus on a letter and when 267 00:11:15,450 --> 00:11:17,519 it lights up, 268 00:11:17,520 --> 00:11:19,619 that is a real and relevant 269 00:11:19,620 --> 00:11:20,620 event to you 270 00:11:22,020 --> 00:11:23,669 because you want to type it. 271 00:11:23,670 --> 00:11:25,829 So exactly in that case, you 272 00:11:25,830 --> 00:11:26,830 will have 273 00:11:28,290 --> 00:11:31,589 the P300 in your ERP. 274 00:11:31,590 --> 00:11:34,349 And of course, 275 00:11:34,350 --> 00:11:36,539 the system that is providing the stimuli 276 00:11:38,010 --> 00:11:40,049 as recordings, the timing and then north 277 00:11:40,050 --> 00:11:42,989 which which letter actually 278 00:11:42,990 --> 00:11:45,629 was lit up when this 279 00:11:45,630 --> 00:11:46,529 appeared. 280 00:11:46,530 --> 00:11:48,809 So this way you can Typekit things. 281 00:11:49,890 --> 00:11:52,079 But again, you would need tool to 282 00:11:52,080 --> 00:11:54,179 stare at one letter for a while 283 00:11:54,180 --> 00:11:55,950 because it has to be repeated 284 00:11:56,970 --> 00:11:57,970 several times. 285 00:12:01,320 --> 00:12:03,449 So I want to present 286 00:12:05,010 --> 00:12:07,230 more specialized piece 100 based 287 00:12:08,700 --> 00:12:11,069 brain computer interface that 288 00:12:11,070 --> 00:12:13,469 is an authentication scheme. 289 00:12:13,470 --> 00:12:16,949 The idea is we are having 100 290 00:12:16,950 --> 00:12:19,109 normal photos can be anything 291 00:12:20,280 --> 00:12:22,679 and we select some which 292 00:12:22,680 --> 00:12:24,000 are all passport. 293 00:12:25,020 --> 00:12:27,239 So this is the example that actually 294 00:12:27,240 --> 00:12:28,919 has had that use for the experiment. 295 00:12:32,400 --> 00:12:34,799 Five very different photos. 296 00:12:34,800 --> 00:12:36,989 So now I'm doing a small experiment was 297 00:12:36,990 --> 00:12:39,480 you try to 298 00:12:41,760 --> 00:12:44,549 remember those photos and 299 00:12:44,550 --> 00:12:46,769 try to see them 300 00:12:46,770 --> 00:12:49,110 and count them in 301 00:12:50,130 --> 00:12:51,330 this video stream. 302 00:13:12,880 --> 00:13:15,069 So I think I'm not 303 00:13:15,070 --> 00:13:16,689 asking you to raise your arms because I 304 00:13:16,690 --> 00:13:18,219 come to you anyway. 305 00:13:18,220 --> 00:13:20,469 I guess some of you 306 00:13:20,470 --> 00:13:22,329 might have seen all five pictures. 307 00:13:22,330 --> 00:13:24,279 My son might have counted less 308 00:13:25,900 --> 00:13:27,459 for the first time. The task is not 309 00:13:27,460 --> 00:13:29,889 really easy. But generally 310 00:13:29,890 --> 00:13:32,409 when in this experiment, you you 311 00:13:32,410 --> 00:13:34,539 counted something, you would probably 312 00:13:34,540 --> 00:13:36,850 have had this piece thrown in your brain. 313 00:13:43,580 --> 00:13:45,760 Those are the results of the experiment. 314 00:13:49,540 --> 00:13:51,669 I hope you can read it 315 00:13:51,670 --> 00:13:52,869 might be a bit too dark 316 00:13:54,460 --> 00:13:56,589 as we have 95 non-target 317 00:13:56,590 --> 00:13:58,899 image and only five percent. 318 00:13:58,900 --> 00:14:01,390 So five target images. 319 00:14:02,560 --> 00:14:04,779 We use the F1 320 00:14:04,780 --> 00:14:07,419 score to evaluate 321 00:14:07,420 --> 00:14:08,420 the classifier. 322 00:14:10,810 --> 00:14:14,109 We did cross validation on the data here 323 00:14:14,110 --> 00:14:15,789 where we have the best score. 324 00:14:15,790 --> 00:14:16,840 We also did 325 00:14:17,950 --> 00:14:20,169 train the classifier, which was a 326 00:14:20,170 --> 00:14:22,599 simple, linear discriminant 327 00:14:22,600 --> 00:14:23,600 analysis. 328 00:14:26,230 --> 00:14:29,019 We trained just by 329 00:14:29,020 --> 00:14:31,209 the experiment done by one person, 330 00:14:31,210 --> 00:14:33,309 and then we also tried to 331 00:14:33,310 --> 00:14:35,469 do a general classifier that 332 00:14:35,470 --> 00:14:37,839 we trained with other people's 333 00:14:37,840 --> 00:14:40,479 data. For that, we used the best datasets 334 00:14:40,480 --> 00:14:43,029 that we have that had the highest 335 00:14:43,030 --> 00:14:44,669 score and the cross validation. 336 00:14:46,510 --> 00:14:48,009 It still works. 337 00:14:48,010 --> 00:14:50,829 The interesting thing here is that 338 00:14:50,830 --> 00:14:53,439 the classification of the IG data 339 00:14:53,440 --> 00:14:56,049 as possible without tuning the classifier 340 00:14:56,050 --> 00:14:58,269 to the user 341 00:14:58,270 --> 00:15:00,279 making the system non biometric. 342 00:15:04,220 --> 00:15:06,399 This is the number of trials 343 00:15:06,400 --> 00:15:09,219 that we average, so we actually scored 344 00:15:09,220 --> 00:15:11,589 50 of those bursts that you have seen 345 00:15:11,590 --> 00:15:13,239 just before. 346 00:15:13,240 --> 00:15:15,309 That takes about 20 minutes, 347 00:15:15,310 --> 00:15:17,499 and no one wants to use 348 00:15:17,500 --> 00:15:19,959 an authentication system where one 349 00:15:19,960 --> 00:15:21,639 lock in takes 20 minutes. 350 00:15:21,640 --> 00:15:23,869 So we looked at how many 351 00:15:23,870 --> 00:15:25,659 birds do we actually need. 352 00:15:25,660 --> 00:15:28,029 But it seems like up to 50. 353 00:15:28,030 --> 00:15:30,189 It still increases, and there 354 00:15:30,190 --> 00:15:32,589 is a huge difference between different 355 00:15:32,590 --> 00:15:34,989 sets or the highest line, 356 00:15:34,990 --> 00:15:37,479 as is the top rated 357 00:15:37,480 --> 00:15:39,969 subjects and zeros are 358 00:15:39,970 --> 00:15:42,309 much lower. So it depends 359 00:15:42,310 --> 00:15:43,719 a lot on the subject. 360 00:15:43,720 --> 00:15:46,659 Maybe also on how was 361 00:15:46,660 --> 00:15:48,909 EEG headset fit and of 362 00:15:48,910 --> 00:15:51,160 maybe how was I focused on the task? 363 00:15:52,990 --> 00:15:53,990 Um 364 00:15:55,630 --> 00:15:57,789 yeah. But it's the 365 00:15:57,790 --> 00:16:00,009 biggest effect that we found was, 366 00:16:01,450 --> 00:16:04,119 um, for an authentication 367 00:16:04,120 --> 00:16:06,569 system, we want to have permanence. 368 00:16:06,570 --> 00:16:08,649 So if we want to lock in again 369 00:16:08,650 --> 00:16:11,079 after a few months, it should still work. 370 00:16:11,080 --> 00:16:13,359 So we had three sessions 371 00:16:13,360 --> 00:16:16,119 with some months in between. 372 00:16:16,120 --> 00:16:18,159 And actually, we got better schools over 373 00:16:18,160 --> 00:16:19,160 time. 374 00:16:19,960 --> 00:16:22,059 We feared that they might degrade, but it 375 00:16:22,060 --> 00:16:23,530 seems there a training effect. 376 00:16:24,820 --> 00:16:27,009 Um yeah. And the signal is permanent 377 00:16:27,010 --> 00:16:28,010 enough. 378 00:16:28,630 --> 00:16:30,519 So here is the final score. 379 00:16:30,520 --> 00:16:31,419 That's a plot. 380 00:16:31,420 --> 00:16:33,249 Because even though it's no real 381 00:16:33,250 --> 00:16:35,860 biometric system, um, 382 00:16:36,970 --> 00:16:39,429 we are measuring biosignatures 383 00:16:39,430 --> 00:16:41,379 and that can go wrong. 384 00:16:41,380 --> 00:16:43,539 Therefore, we have false acceptance rates 385 00:16:43,540 --> 00:16:45,939 and false rejection rates, 386 00:16:45,940 --> 00:16:47,769 and we are coming to an equal error rate 387 00:16:47,770 --> 00:16:49,959 of about 10 percent, which is, of 388 00:16:49,960 --> 00:16:51,729 course, too bad, especially when you'll 389 00:16:51,730 --> 00:16:53,859 see that one authentication 390 00:16:53,860 --> 00:16:56,419 round takes about 20 minutes. 391 00:16:56,420 --> 00:16:57,420 Um 392 00:16:58,480 --> 00:17:00,609 yeah. So the idea was if 393 00:17:00,610 --> 00:17:03,099 we have the we are showing 394 00:17:03,100 --> 00:17:05,348 the images very fast, we can 395 00:17:05,349 --> 00:17:07,059 have a very sort 396 00:17:08,349 --> 00:17:10,029 of lock in time. 397 00:17:10,030 --> 00:17:12,219 But it actually didn't really work well 398 00:17:12,220 --> 00:17:13,568 enough. 399 00:17:13,569 --> 00:17:14,569 And 400 00:17:15,700 --> 00:17:16,700 that's it. 401 00:17:18,010 --> 00:17:19,749 There are three minutes left. 402 00:17:19,750 --> 00:17:20,829 Are there questions? 403 00:17:31,210 --> 00:17:33,969 Oh, I found my voice again, so 404 00:17:33,970 --> 00:17:36,039 indeed, we have a bit of time. 405 00:17:36,040 --> 00:17:38,229 There are two microphones and the 406 00:17:38,230 --> 00:17:40,389 internets, so remember and 407 00:17:40,390 --> 00:17:41,740 that's a question at that microphone. 408 00:17:42,850 --> 00:17:43,850 Hello, there. 409 00:17:44,950 --> 00:17:47,979 I was wondering, have you heard of 410 00:17:47,980 --> 00:17:50,410 this being used to detect terrorists? 411 00:17:51,850 --> 00:17:54,129 There was a there is an experiment done 412 00:17:54,130 --> 00:17:56,469 where they showed images of things 413 00:17:56,470 --> 00:17:58,389 that you're not supposed to know as as a 414 00:17:58,390 --> 00:17:59,769 regular citizen. 415 00:17:59,770 --> 00:18:01,329 So they would show all these innocent 416 00:18:01,330 --> 00:18:03,639 images. And then there would also be 417 00:18:03,640 --> 00:18:05,859 a blasting cap or the magazine 418 00:18:05,860 --> 00:18:08,109 of an AK 47 or stuff 419 00:18:08,110 --> 00:18:09,519 that you're supposed to know if you went 420 00:18:09,520 --> 00:18:11,589 to a terrorist training camp and 421 00:18:11,590 --> 00:18:13,899 they would do the exact same P300 422 00:18:13,900 --> 00:18:16,569 thing. I thought that was interesting. 423 00:18:16,570 --> 00:18:18,999 And of course, if you read about, 424 00:18:19,000 --> 00:18:20,409 you know, about training camps and 425 00:18:20,410 --> 00:18:22,539 terrorists, you would fail the tests, 426 00:18:22,540 --> 00:18:23,859 which would be interesting. 427 00:18:23,860 --> 00:18:26,199 Yeah, um, I 428 00:18:26,200 --> 00:18:27,759 haven't heard about the application on 429 00:18:27,760 --> 00:18:29,889 terrorism, but very similar to the 430 00:18:29,890 --> 00:18:31,900 application on criminal 431 00:18:32,980 --> 00:18:35,169 investigations as a lie detector based 432 00:18:35,170 --> 00:18:36,159 on Piece 100. 433 00:18:36,160 --> 00:18:38,299 So basically the same, 434 00:18:38,300 --> 00:18:40,089 you are showing some pictures that only 435 00:18:40,090 --> 00:18:42,159 the only 436 00:18:42,160 --> 00:18:44,469 the guilty 437 00:18:44,470 --> 00:18:45,399 knowledge tests. 438 00:18:45,400 --> 00:18:47,529 Yeah. Yeah, that that only the one who 439 00:18:47,530 --> 00:18:48,999 is guilty knows. 440 00:18:49,000 --> 00:18:51,249 And but there are also 441 00:18:51,250 --> 00:18:54,129 some papers about if you are 442 00:18:54,130 --> 00:18:56,349 if you are attacked by this piece for 100 443 00:18:56,350 --> 00:18:59,259 based guilty knowledge test, how to 444 00:18:59,260 --> 00:19:00,910 how to prevent being detected. 445 00:19:02,740 --> 00:19:04,959 OK, now the question that microphone 446 00:19:04,960 --> 00:19:05,139 I 447 00:19:05,140 --> 00:19:06,609 have two small questions. 448 00:19:06,610 --> 00:19:08,669 One is which was the EEG headset 449 00:19:08,670 --> 00:19:10,659 used in the image based authentication. 450 00:19:10,660 --> 00:19:12,039 Was it the open BCI? 451 00:19:12,040 --> 00:19:14,109 And then also is P300 452 00:19:16,000 --> 00:19:17,829 like individualize like, does it need a 453 00:19:17,830 --> 00:19:19,659 lot of calibration or is it something 454 00:19:19,660 --> 00:19:21,339 that you can just detect leaks straight 455 00:19:21,340 --> 00:19:22,340 in? 456 00:19:23,010 --> 00:19:25,179 So the eggheads had used 457 00:19:25,180 --> 00:19:27,279 for for the experiments was the emotive 458 00:19:27,280 --> 00:19:28,280 epoch, 459 00:19:29,890 --> 00:19:32,169 and the 460 00:19:32,170 --> 00:19:35,909 P300 seems to be a bit individual. 461 00:19:35,910 --> 00:19:38,499 There are approaches doing biometrics 462 00:19:38,500 --> 00:19:41,559 by looking at P300. 463 00:19:41,560 --> 00:19:43,759 But our approach was to have a general 464 00:19:43,760 --> 00:19:46,029 P300 detector that would work 465 00:19:46,030 --> 00:19:47,649 on anyone. That was the difference 466 00:19:47,650 --> 00:19:50,019 between the ICC individual classifier 467 00:19:50,020 --> 00:19:51,020 and a 468 00:19:52,150 --> 00:19:53,410 general classifier. 469 00:19:54,490 --> 00:19:56,219 You can see the score difference here, 470 00:19:56,220 --> 00:19:58,149 the middle and the right one, and the 471 00:19:58,150 --> 00:19:59,680 left four was a cross validation. 472 00:20:02,410 --> 00:20:03,909 OK, is there a question from the 473 00:20:03,910 --> 00:20:05,629 internet? I'm looking at the no question 474 00:20:05,630 --> 00:20:06,630 from the internet. 475 00:20:07,390 --> 00:20:09,519 Um, so unfortunately, the time is 476 00:20:09,520 --> 00:20:11,799 running out, and so I have to ask 477 00:20:11,800 --> 00:20:13,719 you to approach security directly. 478 00:20:13,720 --> 00:20:15,609 And I don't know. 479 00:20:15,610 --> 00:20:17,589 You're here, you're at the Congress. 480 00:20:17,590 --> 00:20:20,619 Yes, can be contacted in some way, 481 00:20:20,620 --> 00:20:21,519 I guess. 482 00:20:21,520 --> 00:20:23,709 OK, then I would say thanks 483 00:20:23,710 --> 00:20:24,710 again for his time.