I really can't escape the thought that it's just pumping out a generic face of a particular race, though. I'd like to see the faces that it produces for a number of different-looking people of the same race. (The two faces in the linked article certainly look very much like "generic black man" and "generic black woman").
That’s interesting, because the perp must have been looking for reasons that it looks like him, and found them, whereas I ask “look like anybody know? Nope.” even if the kid lived next door. If he’d just kept his mouth shut...
Of course, I’ve always questioned police artist sketches, too. Almost all of them look like every/none white/Black/Hispanic man/woman I’ve ever known. There might be similarities to someone I know, but not enough that I’d have SWAT kicking down their door.
It looks like this is based algorithmically on genetic ancestry. I'd like to see a deep learning model trained from facial image and genetic data to build a tool for generating faces from genes. Anyone know if something like this is feasible? Is it possible to extract only genetic data that has an impact on facial appearance? Or is this something that we still don't know much about?
Maybe somebody is already using deep learning to train DNA with faces. Not sure how the architecture would look like but I think some form of unsupervised learning to find features mixed with GAN.
Let's get to the heart of what this is: a crude pseudoscience.
While DNA call tell you a few obvious traits, like sex and ancestry with high confidence, almost everything else is speculative (i.e. they are LIKELY to have such and such other traits, but in reality they may not have any of those traits at all).
You know what would be interesting, though probably also flawed? Looking at mRNA.
The reason DNA wouldn't work that well is because a lot of DNA we have isn't actually coded into the traits we present. I think this is what happens at a fairly abstract level:
DNA -> RNA(whole) -> Splicing out of non coding (introns?) portions of the RNA -> RNA(coding) -> some more pre-processing -> mRNA -> Proteins
Those proteins are most likely what help us look the way we do. So going a step back, mRNA could be an interesting target for understanding what genes are actually on and coding how we look.
An obvious flaw from the get go: mRNA is extremely difficult to extract from human tissue. Yields are super low in laboratories, and is consistently an issue in molecular biology research. And you'd have to design a primer to amplify the amount you obtain, that matches up with every strand of mRNA necessary to get a good idea of what someone would look like. Not to mention there's probably some step I'm missing in this process that would affect the role of mRNA. Something like microRNAs, maybe?
What makes you think that with a large library of genomic data (only recently available) and a large corresponding library of quantified facial measurements you wouldn't be able to build a useful predictive model?
I'm sure it's not perfect but I wouldn't be so quick to say it's impossible. (The science is being developed by Craig Venter, who may be a jerk but he's often been a jerk at interesting frontiers in science.)
In support of your argument though, I think it would very hard to build a model that was precise enough to distinguish two people who already look alike.
(You probably couldn't use these generated facial predictions to pull a single individual out of a database of photos.)
15 comments
[ 5.0 ms ] story [ 48.9 ms ] threadhttps://www.forensicsciencedegree.org/wp-content/uploads/201...
I really can't escape the thought that it's just pumping out a generic face of a particular race, though. I'd like to see the faces that it produces for a number of different-looking people of the same race. (The two faces in the linked article certainly look very much like "generic black man" and "generic black woman").
[1] https://nypost.com/2017/11/17/man-confesses-to-murder-after-...
Of course, I’ve always questioned police artist sketches, too. Almost all of them look like every/none white/Black/Hispanic man/woman I’ve ever known. There might be similarities to someone I know, but not enough that I’d have SWAT kicking down their door.
https://www.economist.com/news/science-and-technology/217286...
http://deweyhagborg.com/
https://www.smithsonianmag.com/science-nature/creepy-or-cool...
While DNA call tell you a few obvious traits, like sex and ancestry with high confidence, almost everything else is speculative (i.e. they are LIKELY to have such and such other traits, but in reality they may not have any of those traits at all).
This is 21st century phrenology.
The reason DNA wouldn't work that well is because a lot of DNA we have isn't actually coded into the traits we present. I think this is what happens at a fairly abstract level:
DNA -> RNA(whole) -> Splicing out of non coding (introns?) portions of the RNA -> RNA(coding) -> some more pre-processing -> mRNA -> Proteins
Those proteins are most likely what help us look the way we do. So going a step back, mRNA could be an interesting target for understanding what genes are actually on and coding how we look.
An obvious flaw from the get go: mRNA is extremely difficult to extract from human tissue. Yields are super low in laboratories, and is consistently an issue in molecular biology research. And you'd have to design a primer to amplify the amount you obtain, that matches up with every strand of mRNA necessary to get a good idea of what someone would look like. Not to mention there's probably some step I'm missing in this process that would affect the role of mRNA. Something like microRNAs, maybe?
What makes you think that with a large library of genomic data (only recently available) and a large corresponding library of quantified facial measurements you wouldn't be able to build a useful predictive model?
I'm sure it's not perfect but I wouldn't be so quick to say it's impossible. (The science is being developed by Craig Venter, who may be a jerk but he's often been a jerk at interesting frontiers in science.)
https://www.economist.com/news/science-and-technology/217286...
https://www.technologyreview.com/s/608813/does-your-genome-p...
In support of your argument though, I think it would very hard to build a model that was precise enough to distinguish two people who already look alike.
(You probably couldn't use these generated facial predictions to pull a single individual out of a database of photos.)