WEBVTT - I Know That Face 0:00:00.160 --> 0:00:07.200 Brought to you by Toyota. Let's go places. Welcome to 0:00:07.400 --> 0:00:15.320 Forward Thinking. Hey there, and welcome to Forward Thinking the podcast. 0:00:15.360 --> 0:00:17.239 Then it looks at the future and says, then I 0:00:17.280 --> 0:00:20.560 saw her face. Now I'm a believer. I'm Jonathan Strickland, 0:00:20.600 --> 0:00:24.479 I'm Lauren Boa, and I'm Joe mcformick. So I saw 0:00:24.560 --> 0:00:29.720 something absolutely delightful the other day on the internet. You 0:00:29.760 --> 0:00:31.760 shared it with us, yes, and then you thought this 0:00:31.800 --> 0:00:33.320 could be a good blog post, and I said, to 0:00:33.360 --> 0:00:36.600 heck with that, that could be a great podcast. That 0:00:36.760 --> 0:00:39.199 is right, That is what you said, because you know 0:00:39.280 --> 0:00:42.080 a good podcast when you see one, which is odd 0:00:42.080 --> 0:00:44.400 because it's an auditory kind of thing that we do. 0:00:44.440 --> 0:00:47.120 But anyway, well thanks a pun. That would have worked 0:00:47.159 --> 0:00:50.120 if we'd already announced what today's podcast was about. What 0:00:50.200 --> 0:00:53.400 we're about to get to it. Um. So I found 0:00:53.440 --> 0:00:58.600 it was through Alexis Madrigals Five Intriguing Things newsletter, which 0:00:58.600 --> 0:01:01.000 if you're not subscribed to that, you should check it out. 0:01:01.080 --> 0:01:05.600 It's excellent. Um. The link was to a project of 0:01:05.680 --> 0:01:09.399 a Korean artist group that's I believe made up of 0:01:09.440 --> 0:01:14.040 two people, Shin Sung Bak and Kim Yong hun, and 0:01:14.120 --> 0:01:19.080 the project was called cat or human. Okay, So for 0:01:19.120 --> 0:01:21.840 those who are listening who have not had the benefit 0:01:21.920 --> 0:01:25.720 of seeing this particular project, what do you how would 0:01:25.720 --> 0:01:28.520 you describe what cat or human is? First of all, 0:01:29.760 --> 0:01:38.600 it's not horrifying, No, it's it wasn't like a taste test. No. 0:01:38.720 --> 0:01:41.720 It was also not one of those animal human hybrids 0:01:41.720 --> 0:01:43.640 that I keep hearing about from the stuff They don't 0:01:43.640 --> 0:01:47.920 want you to any dr Moreau, none of that. No, 0:01:48.280 --> 0:01:52.600 it is merely a collection of fun of images from 0:01:52.640 --> 0:01:56.680 two different algorithms, and it's marking where two different algorithms 0:01:56.720 --> 0:02:00.200 went wrong. One of them was an algorithm that was 0:02:00.280 --> 0:02:04.720 created to recognize human faces and say that's a human 0:02:04.760 --> 0:02:07.640 face just by looking at a picture. Another was an 0:02:07.680 --> 0:02:11.160 algorithm that was created to recognize cat faces by looking 0:02:11.200 --> 0:02:13.640 at a picture and saying, hey, that's a cat face. 0:02:14.080 --> 0:02:18.560 Of course, what's great is when they mess up, specifically 0:02:18.760 --> 0:02:22.919 when the cat recognizing algorithm recognizes a human face as 0:02:22.919 --> 0:02:26.600 a cat face, right, and when the human recognizing algorithm 0:02:26.639 --> 0:02:30.000 recognizes a cat face as a human face. So they 0:02:30.040 --> 0:02:34.200 collected images from both of these sets of mistakes and said, 0:02:34.320 --> 0:02:38.000 according to these various algorithms, these people are cats and 0:02:38.040 --> 0:02:40.959 these cats are people. That's exactly right. It's kind of 0:02:40.960 --> 0:02:45.680 like a weird yearbook. Yeah, so that is what they 0:02:45.720 --> 0:02:47.959 did from what I can tell that. You should check 0:02:47.960 --> 0:02:51.640 it out on their website. The the cat people, you know, 0:02:51.840 --> 0:02:54.800 look like cat people. I agree. I mean it's I 0:02:54.840 --> 0:02:57.519 guess a little bit. They mostly look like people. Yeah, 0:02:57.560 --> 0:03:00.720 they may have some features that are slightly feline. When 0:03:00.720 --> 0:03:03.320 we say cat people, we mean people who may have 0:03:03.480 --> 0:03:06.200 certain feline features. We don't mean people who have twenty 0:03:06.240 --> 0:03:09.200 seven cats. No, not that thing, and also not the 0:03:09.200 --> 0:03:12.800 thing from like doctor Who, not those cat people either. Um, 0:03:13.480 --> 0:03:16.520 maybe slightly cat like eyes. Yeah, we're not talking to sleepwalk. 0:03:16.560 --> 0:03:18.960 They tend to be very happy looking people with like 0:03:19.040 --> 0:03:23.160 big grins ches cheshire people, which is funny because cats 0:03:23.160 --> 0:03:27.360 are so evil. But anyway, the selected cats I'm looking 0:03:27.400 --> 0:03:30.600 at now that we're yeah, the cats that were identified 0:03:30.600 --> 0:03:34.560 as human, they all look really sour. They're not smiling cats. 0:03:35.280 --> 0:03:37.840 I don't know what like, So grumpy cat might have 0:03:37.960 --> 0:03:41.000 ended up being classified as a grumpy cat is not 0:03:41.120 --> 0:03:43.480 one of the ones I've seen. Instead, these are cats 0:03:43.680 --> 0:03:46.720 usually that are kind of wide eyed and looking straight 0:03:46.760 --> 0:03:49.119 on at the camera and in this kind of like 0:03:49.320 --> 0:03:54.040 what what so the reason why we even bring this up, 0:03:54.080 --> 0:03:57.040 as as amazing as this particular art project is, it 0:03:57.120 --> 0:03:59.760 does have to do with the technology of the future, 0:03:59.840 --> 0:04:03.600 we promise. Yeah, we wanted to talk about facial recognition technology. 0:04:03.680 --> 0:04:06.800 What that has to do with artificial intelligence? What is 0:04:06.840 --> 0:04:09.320 it actually doing, how is it doing that, and what 0:04:09.480 --> 0:04:12.360 could that possibly mean to us? Not to mention some 0:04:12.920 --> 0:04:16.480 interesting looks at the future and apparently at least one 0:04:16.560 --> 0:04:18.719 segment that will scare the pants off me, but I 0:04:18.720 --> 0:04:22.159 don't know what it is yet. Yeah, So this we're 0:04:22.240 --> 0:04:24.120 we're going to go on a journey of discovery and 0:04:24.160 --> 0:04:26.039 some of us are going to be discovering right along 0:04:26.080 --> 0:04:30.040 with you. So to jump into this, Uh, facial recognition 0:04:30.120 --> 0:04:33.360 is part of artificial intelligence. When we talk about AI. 0:04:33.880 --> 0:04:36.000 We've said this several times on this show. AI is 0:04:36.040 --> 0:04:39.560 actually a very broad term, right. We often think of 0:04:39.600 --> 0:04:42.359 it in terms of strong AI, some sort of computer 0:04:42.440 --> 0:04:46.040 that can quote unquote think like a person. But in reality, 0:04:46.120 --> 0:04:52.039 artificial intelligence incorporates lots of different elements of intelligence, right, right, 0:04:52.080 --> 0:04:55.240 And one of those elements would be making sense of 0:04:55.279 --> 0:04:58.440 the kind of data that makes sense to animals and 0:04:58.520 --> 0:05:01.719 humans but not to computer. Right, So, like visual data 0:05:01.839 --> 0:05:04.520 I can look at Joe and I can recognize that 0:05:04.520 --> 0:05:07.599 that's one, that's a person. Two, I can recognize that 0:05:07.720 --> 0:05:10.359 it's a male person. Three, I can recognize that it 0:05:10.600 --> 0:05:14.719 is a person I already know for that that person's 0:05:14.760 --> 0:05:16.440 name is Joe. Like all of this kind of stuff 0:05:16.839 --> 0:05:20.039 a batman costume right now, right exactly not right now, 0:05:20.120 --> 0:05:22.360 that you're not a cat, Yeah, they're there are all 0:05:22.400 --> 0:05:25.240 these things. And I can also have all these other 0:05:25.279 --> 0:05:29.880 associations with Joe that that you know, factor into who 0:05:29.920 --> 0:05:32.400 he is and what he is, and all of this 0:05:32.480 --> 0:05:34.680 is very natural to me. But all stuff we have 0:05:34.720 --> 0:05:37.520 to teach computers right right, and it's very natural to 0:05:37.600 --> 0:05:40.080 humans to be able to I mean, you could probably 0:05:40.120 --> 0:05:42.400 recognize Joe from across the room even if he were 0:05:42.400 --> 0:05:45.039 turned around, due to you knowing him well enough to 0:05:45.040 --> 0:05:47.200 know like his gait and posture and like if you 0:05:47.240 --> 0:05:49.039 catch kind of a weird angle from the side of 0:05:49.040 --> 0:05:52.320 his face, probably still identify that as Joe. But a 0:05:52.360 --> 0:05:55.760 computer doesn't have that natural capacity. Yeah, if I look 0:05:55.800 --> 0:05:58.200 at if I if I meet Joe for the first 0:05:58.200 --> 0:05:59.599 time and he looks me right in the eyes, we 0:05:59.600 --> 0:06:01.920 shake ends, and then the next time I see him 0:06:02.279 --> 0:06:05.400 he's in profile, I'm still able to recognize him. But 0:06:05.520 --> 0:06:09.679 for computers that's not a trivial task, right Yeah, because 0:06:09.680 --> 0:06:13.839 they're they're relying on a way of analyzing visual data 0:06:13.920 --> 0:06:16.440 and breaking that down into data points. That makes sense. 0:06:16.800 --> 0:06:20.560 So it is a big part of artificial intelligence, and 0:06:20.640 --> 0:06:24.280 not just facial recognition, but object recognition in general. Now, 0:06:24.320 --> 0:06:26.960 to get into what facial recognition is doing, there are 0:06:27.920 --> 0:06:29.880 a couple of things we have to factor. One that 0:06:29.960 --> 0:06:31.760 has to be able to recognize what is a face? 0:06:31.839 --> 0:06:34.200 What are the elements that make up a typical face? 0:06:34.800 --> 0:06:37.520 And that ends up being sort of the baseline, all right. 0:06:37.560 --> 0:06:40.400 Beyond that, when we talk about facial recognition, we're usually 0:06:40.400 --> 0:06:44.400 referring to two different things that are common to biometrics, 0:06:44.440 --> 0:06:49.440 whether it's facial recognition, fingerprint technology, hand scans, whatever it 0:06:49.480 --> 0:06:54.040 may be. And that's verification and identification. And these are 0:06:54.040 --> 0:06:58.880 two related but different concepts. So verification might be something 0:06:58.920 --> 0:07:01.560 you should think of with thet afore of a key. Yes, 0:07:02.000 --> 0:07:04.480 verification would be something like if we needed to have 0:07:05.040 --> 0:07:08.880 a biometric system to let us into the office, so 0:07:09.040 --> 0:07:11.960 we might have it might be a two step verification process. 0:07:12.040 --> 0:07:15.160 Right there, there's one step where we might use a 0:07:15.240 --> 0:07:18.840 badge or a pen or something else to identify and 0:07:18.880 --> 0:07:21.520 say this is who I claim to be. Now. At 0:07:21.560 --> 0:07:26.040 that point we would then submit whatever happened. Biometric it's 0:07:26.040 --> 0:07:28.640 asking for, whether it's a fingerprint scan or maybe you're 0:07:28.680 --> 0:07:31.680 just looking into a camera and facial recognition software is 0:07:31.680 --> 0:07:34.160 taking a look at who you are. Well, when you 0:07:34.280 --> 0:07:37.680 created that profile for the system, when you created your pen, 0:07:37.840 --> 0:07:41.520 when you had your fingerprint scanned or whatever it may be. UM, 0:07:41.560 --> 0:07:46.160 then that's when it has uh scanned you and taking 0:07:46.200 --> 0:07:50.000 those data points and associated it with your identity. So 0:07:50.040 --> 0:07:52.800 what it's looking for is a match of the person 0:07:52.880 --> 0:07:56.080 who is standing at the doorway with the person who 0:07:56.160 --> 0:07:58.320 have a profile that it already already has set up 0:07:58.320 --> 0:08:00.520 in the system. Right, Really, this is trying to get 0:08:00.520 --> 0:08:03.360 a computer to do what a human would already do, 0:08:03.480 --> 0:08:06.560 looking through the people of a door. So someone knocks 0:08:06.600 --> 0:08:11.080 on your door says hello, it's Joe, and Jonathan comes 0:08:11.120 --> 0:08:13.320 to the door and looks through the people and sees 0:08:14.120 --> 0:08:18.440 that's not Joe, that's a werewolf. Then I'm denied access. 0:08:18.760 --> 0:08:20.880 But at least you have some other questions you need 0:08:20.920 --> 0:08:24.400 to answer before I will open that door. You know, 0:08:25.040 --> 0:08:29.000 you might have had a rough n hope. Yeah, yeah, 0:08:29.160 --> 0:08:30.760 you might have let your beard go a little bit 0:08:30.880 --> 0:08:34.680 um or or or you know, like Jonathan could also 0:08:34.720 --> 0:08:37.360 look through the people and go, that's Joe. He's wearing 0:08:37.360 --> 0:08:40.719 a jaunty hat. But it's definitely still Joe and let 0:08:40.720 --> 0:08:43.800 you in, I mean, depending on his feelings about johnty hat. Right, Yeah, 0:08:44.040 --> 0:08:45.760 do you take your hat off if you come inside? 0:08:45.840 --> 0:08:48.559 That's Joe and clown makeup, but I still recognize him. 0:08:49.160 --> 0:08:54.880 I'm definitely not come anyway. Yeah. But so that's the 0:08:54.920 --> 0:08:57.640 sort of artificial intelligence angle on it. It's just trying 0:08:57.679 --> 0:09:00.800 to recreate what happens in your brain and when you 0:09:00.840 --> 0:09:03.400 look through the peepole at somebody before you let them in. 0:09:03.600 --> 0:09:05.520 And these are the kind of systems, by the way, 0:09:05.600 --> 0:09:08.600 that can can cause problems for people, even when they 0:09:08.640 --> 0:09:12.760 have a legitimate profile set up. I did an interview 0:09:12.800 --> 0:09:16.120 with Dr John Padfield for How Stuff Works where we 0:09:16.160 --> 0:09:18.439 talked about biometrics in general, and he talked about this 0:09:18.480 --> 0:09:21.320 is this is related. It's not facial recognition, but it's related. 0:09:21.600 --> 0:09:24.719 He talked about working at a job where he had 0:09:24.760 --> 0:09:27.040 to put in a pen and have a hand scan 0:09:27.720 --> 0:09:31.160 before he could come into the office, and he normally 0:09:31.200 --> 0:09:34.280 wears a class ring, and one day he left the 0:09:34.360 --> 0:09:36.760 ring off, and so when he went to have his 0:09:36.840 --> 0:09:39.760 hand scanned, his hand had a different readoubt than it 0:09:39.800 --> 0:09:42.520 normally does because he didn't have this enormous lump from 0:09:42.520 --> 0:09:45.280 a class ring on his finger, and it would not 0:09:45.400 --> 0:09:47.120 let him into the building. He had to wait until 0:09:47.120 --> 0:09:49.440 that actual human game and verified that he was who 0:09:49.520 --> 0:09:51.720 he claimed to be before he was let in. Same 0:09:51.760 --> 0:09:54.480 sort of thing can happen with facial recognition software. Okay, 0:09:54.480 --> 0:09:58.120 but there's another thing that facial recognition software might need 0:09:58.160 --> 0:10:00.839 to do, and that's, as you said, identify vacations. So 0:10:01.400 --> 0:10:04.120 unlike looking through a people and saying, oh, okay, is 0:10:04.160 --> 0:10:06.640 the person standing there who they said they were through 0:10:06.679 --> 0:10:09.839 the door, It's more like looking at a person and saying, 0:10:10.160 --> 0:10:13.000 who's that? Yeah, do I recognize this person at all? 0:10:13.480 --> 0:10:16.880 So verification makes it easy for the system because it's 0:10:16.920 --> 0:10:19.160 just looking for a match of who you claim to 0:10:19.160 --> 0:10:22.440 be and who you are. Identification is more tricky. It's 0:10:22.440 --> 0:10:24.760 saying who are you without any of that who you 0:10:24.760 --> 0:10:27.280 claim to be? Stuff? Right? Right? Right? You don't well, 0:10:27.320 --> 0:10:29.880 I mean, you might have a set of vague profiles 0:10:29.920 --> 0:10:33.080 set up. But yeah, so this would be something like 0:10:33.160 --> 0:10:37.240 a system where the police might use an identification system 0:10:37.240 --> 0:10:40.719 in order to figure out the identity of someone who 0:10:40.720 --> 0:10:43.240 has committed a crime and try and match that against 0:10:43.280 --> 0:10:47.520 their entire database of all the photographs that they have stored. 0:10:48.040 --> 0:10:51.600 This is particularly tricky because again, you can run into 0:10:51.640 --> 0:10:55.760 those situations where the image you get during whatever it 0:10:55.840 --> 0:11:00.000 was that happened may not be the same sort of quality, 0:11:00.080 --> 0:11:02.960 the and and condition as the image that you originally 0:11:02.960 --> 0:11:04.960 have in your database. But we'll talk more about that. 0:11:05.000 --> 0:11:06.959 But yeah, I was going to say that's exactly right, 0:11:07.000 --> 0:11:09.640 because one of the things you might be thinking now 0:11:09.760 --> 0:11:13.360 is identification is harder for a lot of reasons. And 0:11:13.559 --> 0:11:16.679 one of the reasons verification is easier is you typically 0:11:16.720 --> 0:11:20.200 when you're verifying someone's identity, can assume that they're being 0:11:20.280 --> 0:11:23.920 cooperative with the scan, that they're looking directly into the camera, 0:11:24.000 --> 0:11:27.040 that you've set up a lighting situation that's conducive to 0:11:27.040 --> 0:11:30.120 to measuring certain bits of their face. Exactly right, They 0:11:30.200 --> 0:11:32.560 want to be verified, and they're getting in the best 0:11:32.640 --> 0:11:36.880 possible scenario for the system to do that identification. You 0:11:36.920 --> 0:11:39.640 don't know they might be willing to let you scan them. 0:11:39.679 --> 0:11:41.920 They might not be willing, they might not be aware. 0:11:42.280 --> 0:11:45.119 You can't depend on them trying to get in position 0:11:45.200 --> 0:11:47.480 and make it easy for the computer or the hardware. 0:11:48.000 --> 0:11:49.600 But all of us will make more sense once we 0:11:49.640 --> 0:11:54.480 have explained how this facial recognition technology works. So what 0:11:54.559 --> 0:11:57.800 are these computer programs looking for now? Basically, what they're 0:11:57.840 --> 0:12:01.160 doing is breaking down faces into a collection of landmarks, 0:12:01.559 --> 0:12:05.160 right of identical identity points I should say not identical points, 0:12:05.160 --> 0:12:09.080 but identity points. So they're looking and they, depending on 0:12:09.120 --> 0:12:11.680 the system, they break down the face in a different 0:12:11.760 --> 0:12:15.319 way and they tend to get obviously more high resolution 0:12:15.559 --> 0:12:18.880 as the technology evolves. But what they do is they 0:12:18.920 --> 0:12:22.920 look for points like the the corners of eyes, the 0:12:22.960 --> 0:12:25.480 width of the nose, the length of the nose, the 0:12:25.600 --> 0:12:28.840 relationship between the nose and the mouth, the relationship between 0:12:28.880 --> 0:12:31.160 the eyes and the mouth. All of these are little 0:12:31.240 --> 0:12:34.520 data points, nodal points that the system can look at. 0:12:35.080 --> 0:12:38.560 And then it's all the relationships between these points that 0:12:38.679 --> 0:12:42.800 make up the identity. Right, So it's not just oh, 0:12:42.880 --> 0:12:46.680 there's Joe, Joe has two eyes. That's not quite specific enough. 0:12:46.880 --> 0:12:49.440 It's talking specifically about things like the width of the 0:12:49.520 --> 0:12:54.160 eyes from one another, the distance between the the the 0:12:54.160 --> 0:12:56.640 inner part of the eye from one to the other, 0:12:57.200 --> 0:13:00.920 or the triangle that the eyes make with the nose. Uh. 0:13:01.200 --> 0:13:05.920 Imagine that, but multiplied by a dozen or or a hundred, 0:13:06.120 --> 0:13:08.720 and you've got all the different little points that make 0:13:08.840 --> 0:13:13.760 up the the landscape fingerprint, if you will, of Joe's face. 0:13:14.640 --> 0:13:17.440 So that when the system looks at another face, it 0:13:17.520 --> 0:13:20.280 compares those data points, and it looks for those data 0:13:20.320 --> 0:13:23.160 points and then compares it against what's in the database already, 0:13:23.440 --> 0:13:26.520 And if there's a match that reaches a certain threshold, 0:13:26.920 --> 0:13:30.440 then the system says that's Joe. Now, if it doesn't 0:13:30.559 --> 0:13:32.800 reach that threshold, it may say it's someone else, or 0:13:32.800 --> 0:13:34.640 it may say I do not know who this person is. 0:13:34.679 --> 0:13:37.600 I don't have a record of this person, so, uh, 0:13:37.640 --> 0:13:39.920 you know, I can't help you. Um. And it's all 0:13:39.960 --> 0:13:43.160 based on statistical probability, right. So we talked about this 0:13:43.280 --> 0:13:45.520 before in the past. IBM S. Watson is the example 0:13:45.559 --> 0:13:49.080 I love to use, where Watson on Jeopardy would only 0:13:49.160 --> 0:13:51.760 give an answer if it met a certain threshold of certainty, 0:13:51.960 --> 0:13:54.480 and if it was below that then Watson wouldn't guess. 0:13:54.840 --> 0:13:58.760 So that was kind of a calculated UH risk on 0:13:58.800 --> 0:14:00.760 the part of IBM S teams saying, well, we don't 0:14:00.760 --> 0:14:06.200 want it to guess incorrectly because there's right. Um, some 0:14:06.200 --> 0:14:08.800 some of these programs will, for example, if it's not 0:14:09.640 --> 0:14:12.880 super sure who a person is, will spit out a 0:14:12.880 --> 0:14:18.480 series of guesses like like maybe for for human identification. Yeah, yeah, 0:14:18.520 --> 0:14:21.040 they're in fact, usually what you'll do if you're talking 0:14:21.040 --> 0:14:25.320 about if you're talking specifically about identification not verification, and 0:14:25.360 --> 0:14:27.880 you're and you're looking at a person and you're trying 0:14:27.920 --> 0:14:30.880 to mind an enormous database. Let's say that there are 0:14:31.040 --> 0:14:33.840 thousands of people that are within this database, because in 0:14:33.960 --> 0:14:38.400 law enforcement databases there can be many thousand people. Then 0:14:38.800 --> 0:14:41.400 you usually have a couple of different passes at this 0:14:41.520 --> 0:14:44.640 where one past will narrow it down to a certain 0:14:44.720 --> 0:14:49.520 group of people that that fit the general UH scan 0:14:49.920 --> 0:14:52.480 that goes on, and then they'll start looking for a 0:14:52.520 --> 0:14:55.960 more in depth comparison to see which ones are the 0:14:56.000 --> 0:14:58.880 most likely to be the same as the target, the 0:14:58.920 --> 0:15:03.360 one that was captured whatever image or video was that 0:15:03.440 --> 0:15:07.160 you're you're analyzing. So uh, they start looking at things 0:15:07.200 --> 0:15:10.480 like the depth of your eye sockets, the width of 0:15:10.520 --> 0:15:13.880 your nose, your cheap bone shape and height, your jawline. 0:15:14.080 --> 0:15:15.920 When I when I read about this kind of stuff, 0:15:16.720 --> 0:15:19.120 it makes me think of the character creation process of 0:15:19.240 --> 0:15:23.600 something like Skyrim, where you get into the minute details 0:15:23.880 --> 0:15:25.920 of what your character looks like. You're like, well, all 0:15:26.000 --> 0:15:29.320 those things are the things that facial recognition software actually 0:15:29.360 --> 0:15:31.880 looks at to say this is the person, because it's 0:15:31.920 --> 0:15:35.200 what makes you look like an individual human being. Aren't 0:15:35.200 --> 0:15:39.960 there some cat people in Skyrim? Yes? There are. Yeah, 0:15:40.000 --> 0:15:42.920 it all comes back. I believe they're what they're called 0:15:43.400 --> 0:15:46.760 head Head shape comes into play too, like the distance 0:15:46.800 --> 0:15:48.880 from the eyes to the ears, the distance from the 0:15:48.920 --> 0:15:51.080 center of the nose to the ears, any facial hair 0:15:51.160 --> 0:15:53.920 that you may or may not have, your hairlines, stuff 0:15:53.960 --> 0:15:56.120 like that, and that's one of the things that also 0:15:56.160 --> 0:15:59.560 can make it tricky for a verification. For example, if 0:15:59.600 --> 0:16:02.880 you have a distinct hairstyle and then you change it, 0:16:03.240 --> 0:16:06.640 then you may encounter a system that no longer really 0:16:06.680 --> 0:16:09.400 recognizes you as easily and it may mean that you 0:16:09.480 --> 0:16:12.600 have to go and update that profile with whatever you know, 0:16:12.640 --> 0:16:14.920 however you've changed. And of course there are other things 0:16:14.920 --> 0:16:18.640 that can cause some issues too, but robust systems usually 0:16:18.680 --> 0:16:21.760 will will depend upon multiple measurements. They're not looking for 0:16:21.840 --> 0:16:24.560 just one thing. Like if you're looking at a very 0:16:24.600 --> 0:16:28.480 simple facial recognition technology, where it's maybe something that's built 0:16:28.480 --> 0:16:30.560 into a digital camera, and all it's supposed to do 0:16:30.640 --> 0:16:33.760 is just detect that there is a face. Period. Yeah, 0:16:33.920 --> 0:16:37.640 it's usually really simple stuff. It's just looking for for uh, 0:16:37.920 --> 0:16:42.200 for shapes that correspond to what it thinks of faces, 0:16:42.440 --> 0:16:45.360 which means they can sometimes detect faces where there are 0:16:45.400 --> 0:16:49.120 no faces. I've seen this happen in nature photography, where 0:16:49.240 --> 0:16:51.640 a person will hold up a camera at a tree 0:16:51.680 --> 0:16:54.600 that has an interesting pattern on it and it will 0:16:54.600 --> 0:16:56.640 detect it as a face. You know, you want to 0:16:57.720 --> 0:17:01.280 exactly and you're thinking, you think, this isn't Game of Thrones. 0:17:01.840 --> 0:17:03.720 The tree does not have red eyes. Now that's not 0:17:03.760 --> 0:17:05.560 a This is not one of those trees that the 0:17:05.600 --> 0:17:09.520 Starks talk about all the time. So there are times 0:17:09.560 --> 0:17:13.440 where that kind of technology can rely on a very 0:17:13.520 --> 0:17:16.640 simplified view of what is a face. But for things 0:17:16.680 --> 0:17:20.560 that are really about verification and identification, they usually depend 0:17:20.640 --> 0:17:24.040 upon multiple inputs to make sure because I mean, clearly, 0:17:24.080 --> 0:17:26.919 if you're talking about verifying someone's identity to give them 0:17:26.960 --> 0:17:30.879 access to something that could be really uh you know, 0:17:31.160 --> 0:17:35.359 classified information or just uh top secret type stuff or 0:17:35.400 --> 0:17:39.280 just just sensitive materials. Then you want to be absolutely 0:17:39.359 --> 0:17:41.360 sure that that is who they say they are. So 0:17:42.440 --> 0:17:44.359 some of them will even go so far as to 0:17:44.440 --> 0:17:47.200 use multiple cameras to get more of a three dimensional 0:17:47.240 --> 0:17:49.920 look at a person, because we know, if you're using 0:17:49.920 --> 0:17:52.480 a single camera with a single lens, it's a flat image. 0:17:53.480 --> 0:17:55.520 You know, it's once you see in a regular movie 0:17:55.640 --> 0:17:59.200 or a regular picture, regular photograph, you don't get any 0:17:59.280 --> 0:18:01.400 real sense of up there. I mean, you can see 0:18:01.400 --> 0:18:03.480 contours and stuff, and you can tell if something is 0:18:03.520 --> 0:18:06.280 closer or further away, and if something has been lit 0:18:06.359 --> 0:18:09.920 really well, it might look three dimensional to you and 0:18:10.160 --> 0:18:12.080 you might even your your brain plays a little bit 0:18:12.080 --> 0:18:14.360 of a trick, but it's still a two D image, right, 0:18:14.400 --> 0:18:17.840 And so if you're using a verification system or any 0:18:17.880 --> 0:18:22.040 identification system that's reliant on a single camera, then that 0:18:22.119 --> 0:18:24.640 makes it a little more complicated too, because it's looking 0:18:24.680 --> 0:18:27.360 at three dimensional images but representing it as a two 0:18:27.400 --> 0:18:32.080 dimensional image. So it that that has some elements of 0:18:32.640 --> 0:18:34.919 difficulty there. If you have multiple cameras, then you can 0:18:34.920 --> 0:18:38.200 get around that. You can rely on multiple inputs to 0:18:38.280 --> 0:18:41.280 figure out how you know, how sure are you that 0:18:41.359 --> 0:18:43.919 this person is who you think they are? Okay, but 0:18:44.040 --> 0:18:47.560 we know facial recognition is not perfect yet. So what 0:18:47.600 --> 0:18:50.919 are some of the challenges that are really facing facial 0:18:50.960 --> 0:18:55.960 recognition software today. Well, facial recognition faces many problems, Joe, 0:18:56.800 --> 0:18:59.960 thank you for throwing it to me like that. Uh, well, 0:19:00.040 --> 0:19:03.400 we talked about the lighting issue. That's a really good example. 0:19:03.520 --> 0:19:08.160 So if I go to create my profile, uh for 0:19:08.240 --> 0:19:10.320 the company I work for. Let's say that I'm working 0:19:10.320 --> 0:19:14.960 for you know, um M six, I've somehow gained clearance 0:19:15.000 --> 0:19:19.680 to work in Britain's Superspy uh Academy. And I don't 0:19:19.720 --> 0:19:21.680 know how I've managed to get there, but I don't 0:19:21.720 --> 0:19:24.119 really doubt that you have. I think that's a pretty 0:19:24.240 --> 0:19:27.760 likely Yes, this accident I'm putting on is actually not 0:19:27.840 --> 0:19:30.600 my real accent. But at any rate, let's say that, uh, 0:19:31.000 --> 0:19:32.800 I need to go and fill out my profile, and 0:19:32.840 --> 0:19:34.680 so they put me in a well lit room, they 0:19:34.680 --> 0:19:37.439 take my photograph from several angles. All of that is 0:19:37.480 --> 0:19:40.680 factored in so that I can get into Superspy Academy 0:19:40.720 --> 0:19:43.719 whenever I need to, uh, And then I walk up 0:19:43.760 --> 0:19:48.800 but Superspy Academy. Unfortunately, maintenance hasn't changed out a lightbulb 0:19:48.840 --> 0:19:51.680 that really needs to change. It's gotten very dim, it's 0:19:51.680 --> 0:19:54.800 not giving nearly enough light, and that that dim light 0:19:54.960 --> 0:19:58.120 is now casting odd shadows on me. And that might 0:19:58.160 --> 0:20:01.760 be enough to throw off a facial recognition system. I 0:20:01.760 --> 0:20:05.720 mean that you're ultimately talking about detecting light that's reflecting 0:20:05.760 --> 0:20:08.600 off of a person. And if the lighting is different enough, 0:20:09.000 --> 0:20:12.680 then from whatever your base photograph is, whatever your base 0:20:12.760 --> 0:20:15.560 amount of data is for your the system, then it 0:20:15.680 --> 0:20:19.240 may not return enough information, or at least not enough, uh, 0:20:19.280 --> 0:20:22.359 you know, a good quality information for the system to 0:20:22.359 --> 0:20:25.080 be able to identify you. This is true whether it's 0:20:25.119 --> 0:20:28.320 a security camera that's on the wall, whether it's something 0:20:28.359 --> 0:20:30.800 that's actually set up to be a verification system. I mean, 0:20:30.840 --> 0:20:34.320 it doesn't really matter what the context is. It's it's 0:20:34.320 --> 0:20:37.960 a real problem. Yeah, how about something even simpler. The 0:20:38.000 --> 0:20:41.600 way you're holding your head? Yeah, yeah, greater than fifteen 0:20:41.640 --> 0:20:45.480 degrees of tilt can really throw a wrench in facial recognition. 0:20:45.640 --> 0:20:48.200 But I've seen this in action. Actually, when I've I've 0:20:48.200 --> 0:20:51.280 looked up pictures of faces that we're supposed to be blurred. 0:20:51.280 --> 0:20:54.040 We'll get into that in a minute. But weren't a 0:20:54.080 --> 0:20:56.560 lot of the times a program that's supposed to detect 0:20:56.640 --> 0:20:59.760 a face and blur it will miss a face that's 0:21:00.000 --> 0:21:02.919 I agonal, right. Yeah. I mean if if you're looking 0:21:03.000 --> 0:21:07.040 for a specific uh series of shapes and that is 0:21:07.080 --> 0:21:10.080 what defines a face, and that those shapes are in 0:21:10.119 --> 0:21:13.640 an orientation that's different from what you would expect, then 0:21:14.440 --> 0:21:17.119 you can understand why a system would overlook that and 0:21:17.160 --> 0:21:20.679 not recognize it as a face. Yeah. So my suggestion is, 0:21:20.720 --> 0:21:23.560 if you're a criminal trying to avoid the law, just 0:21:23.640 --> 0:21:26.919 do the limbo all the time. Just lean over. My 0:21:27.000 --> 0:21:30.480 suggestion is don't go crime doesn't pay. That's my suggestion. 0:21:30.520 --> 0:21:32.840 But Joe, you know, we understand that you're in the 0:21:32.880 --> 0:21:35.480 ferious now. Now that you've let on, I am gonna 0:21:35.520 --> 0:21:39.320 have to train my superspy techniques on you. So just 0:21:39.400 --> 0:21:42.560 fair warning. Uh, but yeah, I mean they're there. We 0:21:42.640 --> 0:21:45.960 have the same thing about the single camera systems having 0:21:46.000 --> 0:21:50.000 problem detecting depth. That means that you might be able 0:21:50.040 --> 0:21:52.720 to fool such a system by holding up a photograph 0:21:52.960 --> 0:21:56.040 of let's let's say that you have captured a superspy 0:21:56.080 --> 0:21:59.360