Identifying a friend or family member from a baby photo seems like a trivial task. Conversely, showing someone a photo of a child and asking them to determine the corresponding adult can be immensely difficult. Why does matching the faces of friends and family to photos create the illusion that it is a simple and straightforward task?

In this episode, we look at hindsight bias. Why do our brains present versions of the past to suit the present and is there a qualitative difference between image recognition and extrapolation? We discuss hypothesis generation, intractable computational problems, and the limits of probability distribution in analysis. Finally, we see what evolutionary insight can be gleaned from matching photos of babies to their adult selves and put our own biases to the test by interpreting family photos.

A few things we mentioned in this podcast:

• 30 celebrity baby photos
• Unfamiliar face matching with photographs of infants and children
• Unfamiliar faces are not faces: Evidence from a matching task
• When age-progressed images are unreliable: The roles of external features and age range
• Comparative evaluation of automatic age progression methodologies

For more information on Aleph Insights visit our website https://alephinsights.com or to get in touch about our podcast email [email protected]

Hello and welcome to the Cognitive Engineering podcast produced by me, Fraser McGruer for Aleph Insights. In this series of podcasts we take a look at interesting topics and discuss what we think they tell us about analysis and decision making. I'm here with Nick Hare and Peter Coghill of Aleph Insights and this week we're discussing baby photos. Peter, why are we discussing baby photos? Well there's a number of reasons. So

my daughter is now nearly a year old so she's getting really kind of active and really quite into things and one of the things she really likes doing is looking at the photos we have of various family members around the house. Oh right, okay. And so she's definitely recognises me and Rosie in the pictures and I think she's starting to recognise grandparents that she sees quite often in photos as well. And now obviously we are taking lots of photos of Ada like kind of nearly every day we've got photos of Ada doing things. So yeah the idea of photoing babies. We're looking at these photos of family members and some of them are so that she has two cousins who are now one of them's 12 or 13 so she's developing into an adult sort of stage now so she's starting to look like what she'll kind of look like for the rest of her life in some way. But there's also baby photos of her when she was less than a year old and it got as we were looking at these photos it struck me that you can sort of you can totally pick out a baby photo of somebody that you know. So if you if you know so if I know what you look like for instance I've seen you several times if you presented me with a sort of array of 50 baby photos one of which was you as a baby yeah it would be trivial looking one it would be trivial to pick out the baby right. But if I didn't know what you look like say you presented me with a baby photo of somebody I didn't know right it's sort of it's unprofound to say but it would be it's impossible to sort of extrapolate what you would what they would look like as an adult and yet that you can go back the other way very very easily you can sort of there's a sort of directionality to this kind of way of looking at yeah your understanding. Now it's not very profound because obviously there are characteristics in you can look for in the adult picture that you'll find in the baby picture and you don't know necessarily what characteristics are going to survive through to to maturity but it just struck me as odd that there's a kind of as a struck me as a good example of the sort of directionality of knowledge of understanding is that like you can pick out people's baby photos but you can't extrapolate from the baby photo to what they all look like yeah that's it yeah that's it as Nick has been playing with with his high-tech algorithms there are various apps and things will let you do this that sort of take a guess at what they sort of artificially age you or make you younger or do various other thing changes to your to a profile picture but yeah I just thought it was interesting so I thought I want I thought we'd explore it and think about what does it tell us about how our brains work or how our understanding is formed. Now I know we have pre-prepared

or just prepared some photos of ourselves as babies but maybe we should sort of you know loop back to that right at the end right? Sure. Or do you want to go straight into it? I well first of all

I do want to question not exactly challenge but question Peter's assertion I'm not sure it's as trivial as it he makes it out to pick out baby photos and in fact I mean studies into the accuracy of people's ability to match faces show it's surprisingly hard for people with unfamiliar faces right so if you're given a face and then shown 10 faces and asked which and this is not even a difference in age right the same person at the same age but taken from a different angle or whatever to say which face is the is the reference image people are really bad at it it's like a 72% accuracy or something when comparing to you know in various studies they've done you know 70 80% accuracy when they're presented with 10 faces and asked to identify which one is the same as a face they're looking at right there and in I think one experiment I read about where they were asking whether two children were the same child even then only 72% accurate and comparison between a baby and a reference image four to five years old was something like 64% accurate it's not bad but it's it's obviously not a trivial matter to just go right that out of these 10 babies that person there is this baby it's not we're not spectacularly good at it certainly a lot better than chance but only about you know let's say sort of you know five six times better than chance

which suggests I don't know I'm slightly lost but also you're using the word trivial to it sounds like the way you use trivial means like easy yeah yeah in other words like oh it's

it's obvious now it is I think when you're shown someone as a baby and particularly if it's someone you know I think it's very easy to retrofit a recognition to that I think it's very easy to go oh yeah look same same hair same ears but in fact uh you know if the hair and ears weren't the same you'd go oh yeah same chin same nose do you sort of mean I think I think I think we may be fooled into thinking this is easier than it is because it's so easy to find similarities

um you know between between features this I mean it suggests to me this is about um what we need to know what we need to understand what we need to to remember maybe it's that and it's still not clear to me whether we've decided that we're good at this or not or even if we're

talking about the same thing that we're measuring I think Peter's point is stands right which is that there's clearly a qualitative difference between uh you know identifying that someone is the same as a baby and extrapolating what a baby person will look like from a picture of them

as a baby yeah and it feels like I think but also I think the experiment you cite is slightly the constrained lab environment I'm talking about people that you know really well so you've got a real you know I've seen Fraser not just face on on the camera but I've seen the back of his head the side of his head you know I've seen I've seen him I've seen him move I've seen him talk so I've got my brains I am real yeah it's got a model for what brain how Fraser's face looks and moves when he when he scratches his chin or whatever so yeah um uh I've got a lot more data there it's not just I have I haven't just been presented by the photo so I can yeah I could totally I could totally see why your your uh accuracy rate would be really low if you just presented with a picture of somebody and said pick out the same person from a set of static pictures um but this is this this is this feels different you know I could I I I've never actually tested this but it feels really feels really easy to pick out oh that's a picture of cousin Freya uh as a baby and that's a grandfather as a baby um from family photos yeah so and if you had like if you took 10 people that

you knew pretty well um but for whatever reason had not known them as children um and then you were asked to match photos I think I think you get close to 10 out of 10 pretty good yeah yeah which so you're right there is a difference between that and the experiment that you quote

yeah but the crux of the matter really is that is there's the the difference in matching versus

extrapolation this yeah exactly and then so reversing that would we be able to probably not

but why would we I mean I don't know it's well I'll tell you so why one one application which is uh where you see this used is to try and find out what missing persons will look like uh when age progressed so so you know and and obviously then that's you're only really going to find out how

accurate that was after after the person's found no but sorry sorry I mean you're right well why would you that is a good example but what I mean is why you know in human evolution why would you

why would we ever need to that yeah well well I mean look paternity I think no really important because of paternity because you know that is you don't know it's fairly easy to tell whether a baby belongs to a mother because they come out of their wombs I don't know if you know this but uh but the paternity is very important so I think recognition of similarity between a father and a child in particular you would expect to be something we have evolved to be quite good at and and I think on on Peter's point there um it's interesting when you mentioned photos and familiarity that actually often the baby won't look like necessarily exactly like the parents but in a way that's incredibly freaky and which um I think only people who actually had small children will understand the mannerisms of a baby can be strikingly similar to those of the parents you know the kind of facial expressions and um things where you think well I haven't had time to learn this you know the way they look when they're surprised or something um and as a case

this won't stand up on radio but uh I've got pictures of me up here uh for me with Ada I see when she pulls a particular face her the face she pulls is exactly like a member of my family um so I've got a picture of me and my grandfather so this is a picture of my grandfather I think it's my aunt's wedding so this was nineteen seven late 70s you're gonna screen share yeah I'll screen share but let's screen share so we can see it won't work that's right we can

we can use words and stuff we can use words yeah we can paint a beautiful Coghill picture

yeah so uh let me uh let me try oh my god let me try and zoom it in so the two are kind of next to each other so can you see this picture of me here yeah right so what I can see

in front of me picture of my hold on hold on hold on so the pictures that we've said first all is a beautiful color photo of a boy of about I don't know three years or so in a kind of floppy hat floppy hat beautiful blondish curly hair very blue eyes got some food or some other grub

underneath my nose uh yeah so nothing changed there nothing changes there yeah so that's okay note the expression and the sort of elongated part under the nose I see it yeah

he said the child is sort of looking a bit smug I think you'd say smug yeah in a way

a sort of smug and that's half smile and now now now on screen a wry a wry grin or a wry

almost yeah almost a smile but not quite a wry quizzical look but what we're looking at now is it's a black and white photo of a man probably in his 60s or so um wearing a nice sort of suit and tie and a carnation and what were you explaining to me there that he's got he's um

Peter I think he's got the identical expression on his face to the one I was got a little right and is this your father no this is my grandfather this is uh my grandfather my mum's dad um and I notice exactly the same expression in Ada exactly the same and that's like for me that's my paternity test it's like she's pulling grandpa's face that's yes that's my child unless your grandfather's

been having it away with your girlfriend anyway that's yeah can't imagine anyway look uh enough

of that yeah excellent yeah no that is a really good really so i think i i think i have we have

got machinery for spotting doing a sort of paternity test based on visual characteristics and that that's my evidence for that is that um i see this face i see i see grandpa bob um bob's face in in me in my old photos of me but also in my daughter yeah awesome so what's our point

where are we well look so just think age progression why is age progression hard but recognition is easy this thing well in practice right so some of the reasons that i've seen cited as to why age progression is is not straightforward and and why you know even though there is a lot of similarity or features people age in predictable ways and features are retained in predictable ways um you know eye color stays the same hair type stays the same skin color stays the same and and you know the way that people age obviously you can the fact that you can imagine an old person and a young person shows you that there are things that change that are kind of predictable um but there are obviously the big the big differences um come the further out you go so the longer the the time frame you're talking about because of the long-term influences of things like your lifestyle you know the sort of diet drugs exposure to the sun those kinds of things um and of course the fact that certain things change at unpredictable times like when you go bald or gray or when you get wrinkles um can vary widely uh and of course then you've got things like you know your weight or your hairstyle which can change over quite a short period of time so those things um all are quite challenging for age progression and um i think you know machine learning approaches uh are based on training training software to essentially learn um you know you train it on a database of images of the same person over time so that they can then um what you're effectively doing is predicting what that would be like for a new face you take a new face of someone who's 10 or 20 and predict what that would look like based on what you've learned from the database um and uh but of course even then i mean what you're what it's trying to do what it's going to deliver is a single image but what we really ought to want is a kind of probability distribution of faces you know what we ought to see is a range of possible faces and that range might be pretty broad so it's problematic because you know well the the demand is for a single image what will this person look like in 10 years time but but in fact really we want to express that as a kind of probability distribution um so anyway that's just in just a practical point about you know age progression why it's why it's hard okay but i think there are there are lots of things that this relates to or feels related to um one thought is um you know extrapolation it's a bit like hypothesis generation you know it's you're you're trying to generate possibilities um when you're extrapolating and um and then potentially test them you know what are all the faces that this person might have in 10 years time and then the second question of you know which one is it which one is most likely or how likely are they to have each of these faces that is a much much harder question than being presented with two images you know an old person young person saying um what's the probability that these these two are the same and and i mean in general hypothesis generation is a very different process you know the kind of creative thought process you have to use to forecast things you know what's going to happen in the middle east what's going to you know what's going to happen to space travel in 50 years time um is a much harder question than if i present you with a particular scenario and say how likely is that and i think this this sort of feels like that distinction we've got the distinction between on one hand trying to work out what someone's face might turn into and the second question which is just simply testing is that you know are these two faces the same person which is a much less challenging task um and it also kind of reminded me a bit of the distinction between p and np and this the problem of whether they're the same in computer science it's not exactly but it feels similar um i hesitate to ask but what the hell does that mean well so there are these kind of problems you know maths problems or problems that can be solved through through or which you can try and solve by by um you know computation by writing computer programs essentially um problems that fall into the p class are soluble relatively quickly um and uh they are things like dividing a number by 10 right so it's quite an easy quite an easy problem um and it doesn't really matter how long that number is you know if you if you if you if the number's a million or a hundred billion it's only a little bit longer that it takes to do the division it doesn't it doesn't expand you know exponentially it expands polynomially um np problems are the ones that can be verified um in in uh polynomial time so quickly so so np problems that take p problems are the ones where you can find a solution quickly np problems you you so for example if i give you a really big number that's that's the product of two primes it's quite slow to try and find those primes but if i give you the primes and i say that's them you can check it really really quickly you can just go you can just times them together and that's a really quick uh process and you can say yes this this is you know this is the product of these two numbers but trying to find those numbers is very slow because you have to brute force your way through all the possibilities um uh and the big unsolved problem is whether actually p problems are the same as np problems or whether they're different and and if if it turns out that um np problems are in fact the same as p problems then all of these problems all of them become easily soluble and and you can prove that were we to find that p and np were the same all of these np problems that are easy to verify would also have an easy solution having said that this is really not my area at all that's just my kind of under high level understanding of it um if anyone out there thinks they know the answer well the uh clay foundation i think is going to give you a million dollars so write it down send it in two things not two things not to us to the clay no no don't send it to us yeah

no send it to us and then we can send it to them and claim it as ours right good point um yeah so look two things one i have no idea what you're talking about i i've no yeah i don't understand one two how the hell did we get from this mona lisa-esque sort of um quizzical smirk or whatever um rye smile in the in the cog hill generations from that to p and np i do not know i kind of do a bit actually but more salient is we're we're getting towards the end what where are we what are we talking about i don't know i've kind of completely lost where we are in this what do we want to say what get us on track uh peter uh well i well to answer your question i

think nick was suggesting that this problem this this the the easy which you can match to individuals versus the difficulty of extrapolating what somebody would look like

yeah as an analogy was an analogy i do kind of get it i do i do actually i do i do get it so

future will be more predictable than it is yeah no absolutely and indeed we i think we've covered that elsewhere in in most notably in um in some of our prediction podcasts i think i think so with us elections with me making flipping great wadges of cash um okay so anything else we want to talk about before we get on to um our baby photos who wants to go first uh you or me nick

well i'll show so i've got a bit of a progression here and i think you know this thing about babies i honestly i think babies look the same this is me with my with my granny let me try and okay not terribly clear there let's see if i can let's go yeah try closer there's a better one of me i

think possibly even on the same day yeah okay just wait no just keep it there for a minute keep it there so i'm just describing what i'm seeing which is nick is holding up a square photo um colored um looks like from the mid 70s ish maybe a bit later he's wearing this is this young man is the camera is angled down towards him he's on a blanket having a picnic um he looks basically like a skinhead like a ne'er-do-well it's gonna end up baby and no hair but he's got a tartan blanket behind him and a rather natty green um top with a yellow baby girl actually is it and it's probably the nicest thing in the photo actually but yeah keep going well here's one

of me looking distinctly girly at my christening for that you just look like any other baby exactly um but but then when we get a little bit further on i think this is me probably in about um at

the age of about two or three oh my word it's your son now that basically that is your son

oh yes indeed yeah yeah yeah yeah yeah yeah i like that one that's nice there we are and then

and then from then on i think you'd say that that it becomes recognizably me but i think

you'd struggle before that age yeah yeah yeah i don't because i recently moved house i don't particularly have all my photos are sort of bundled around different places but um let's have a look i just had so let's see this is now this is i can't see what i'm holding up right

let me just describe this we're looking at a black and white photo of two well of one very beautiful actually colored but it's very faded but yeah go on and one uh you know slightly slightly alarming looking child with a bowl haircut uh there's the very cute the very cute one who actually i have to say does look a bit it looks very like your your son your twin sons there on the left yeah and the one on the right who looks like he's got a little he looks like he's wearing a shirt and tie he is he looks in fact a bit like boris johnson the one on the right

which one is you fraser so of course i'm the i'm the super cute one of course yeah so um yeah i don't know do you think that looks anything like me peter the the little one uh yeah yeah i'd say

that you're this that you're the cranial shape is definitely you yeah i think you're clutching these straws um and the eyes are very i like your eyes they're big and round so yeah yeah um all

in all gorgeous um but actually just by chance i've got a photo of my daughter on my desk here where i think if you look at those two i you know yeah i think there's a lot similar there especially the lips you know yeah agreed yeah all right um oh well good luck editing this one to be as gripping as unusual efforts so i'm not sure we kind of have to stop but i mean i'm not sure what we've achieved with this podcast um we've had a bit of fun we've had fun it's the friends we made along the way yes exactly yeah that's it i've got nothing to wrap up on i'm too wrapped up with memories of 30 odd years ago now um all right um so we're going to stop there um in terms of the conclusion make your own um and suffice to say unless anything has anything got anyone got anything to add before i close things off i don't think so no peter anything no i think that wry little grin says it all from peter um all right so we'll stop there the cockles we'll stop there as always thanks for listening if you've got any thoughts or suggestions for topics you can email us at podcast at aleph insights.com we'd love to hear from you if you've enjoyed the podcast peter uh what should you do well if you've enjoyed the podcast and

haven't yet subscribed you can find us on apple podcasts or wherever else you get your podcasts

nice very straight i like it yeah yeah yeah thanks as always for listening i'm Fraser McGruer we've been here with Nick Hare and Peter Coghill of Aleph Insights until next time goodbye