I think the 'Phrenology's back, baby' comment in the thread below captures it quite well. Aside from the ethical implications of designing a camera that automatically classifies people as suspicious based on their gait, what the paper appears to actually cover is that an ML process can accurately classify the movement differences between people walking towards a chair to pick up some papers on it and people walking towards a chair to pick up some money under it...
I didn't see any distinction between types of "deception" in the paper which means it's classifying things like social anxiety and generally self-conscious behavior as deceptive. While strictly true it makes the algorithm useless for detecting security threats at this stage.
I can’t believe something like this would be effective against say pickpockets: people for whom an illicit activity has become commonplace in their lives and become natural.
Believing is something I'll do in houses of worship. To learn whether it will be effective against pickpockets, let's wait for (or do) more research instead? I don't find it that far-fetched or unlikely.
Also: what does "we recorded many participants performing tasks involving deceptive walking" mean? Does this AI detect those who perform deceptive walking, or people who are actually deceptive walking?
They had participants carry a folder between two points, one contained papers and one contained money that they were told to conceal and hand off to a person wearing a hoodie and sunglasses. They compared the way the participant walked.
You missed the key part of the experiment: They basically completely changed the experimental setup between the two runs: The confederate changed how they were dressed, the item was on the floor instead of on the table, and, most importantly: they told the participant to hide what they were doing and otherwise act sketchy!
So yeah! No surprise! They told the participant to act sketchy, and they did! What did the participant do? They put their hands in their pocket and looked around a bunch more than normally. Why? Because they were basically told to! That's not "deceptive walking".
Where is the control for picking up the papers off the floor, or the money off the table? How do we know it's not just that people will walk differently when they are about to crouch down? Maybe that would be an interesting paper by itself, but instead the authors think that this type of research is capable of identifying "deception" with 93% accuracy? Absurd!
There is so much wrong with the experiment, that you could write a rebuttal of almost every decision they made.
You're not kidding.
"In this procedure, the deceptive walk is induced by the
experimenter during the briefing, the type of object, and the
appearance of the confederate"
This is not deceptive walking, this is honest walking. They are honestly trying to broadcast the signal that they were told to broadcast.
They annotated video of people wearing hoods, putting hands in pockets, etc, and then trained the classifier to detect people wearing hoods, putting hands in pockets, etc..
The "deceptive" part is just a frosting layer of prejudice they laid on top.
This seems sorta unethical, like this could very easily be biased against one group by training data. Then if this were rolled out to a Target, they could use this to profile people and just point to the "algorithm". Oh, the computer says you look "shady", you can't shop here...
The fact that we don't have a good metric of intelligence doesn't make iq a useful measure of intelligence.
IQ is only useful in determining pathologies, at the lower end, where any test is usually helpful. For anything else, it's almost entirely snake oil.
If you want to know whether someone is good at something, the best way to test that is to do a test for that particular thing. A test for general inteligence, based on the assumption that general intelligence aptitude is normally distributed in the population, is just not good for anything, and is unlikely to be a reflection of reality, it's pure scientism (instead of science).
There are certainly ways to glean information from people that is otherwise unavailable to a naive observer, like looking at how they walk and intuiting their motives. The issue is when others appropriate the language and rhetoric to serve their own selfish goals, e.g., eugenics.
>As of fall 2017, UNC-Chapel Hill had a 7.8 percent enrollment rate for Black students, one of the lowest in the state, while just over 22 percent of residents in North Carolina are Black.
Not to mention, there doesn't seem to have been any control in this experiment. What a bunch of nonsense. These students were told to act sketchy, and predictably probably hammed it up to please the experimenters (see: they stuck their hands in their pockets, looked around a lot, etc).
Wouldn't you want to look at the percentage of black people that apply to the college(who may or may not be from out of state), instead of just looking at the percentage of state residents?
If only 5% of applications to UNC-CH are black, then having 7.8% enrollment is actually good. If 22% of applicants are black, then that is not very good.
There's a lot more to factor in too like the GPA / SATs of the applicants. You don't want to just mirror the entire population regardless of qualifications.
You want to increase outreach to underrepresented groups and adjust for biases but then once you have your applicant pool you probably want your selections to mirror the diversity of the applicant pool rather than the state as a whole.
I am scared that our "AI" (really ML) future will be ML trained from bad data and it's pre-crime-fun for anyone that just acts a bit weird or tries to not stand out too much.
I hope there are enough people like you in the institutions where it matters to be all "hey, wait a fuckin SECOND here".
Yeah, like people with high functioning autism spectrum disorder who need to mask themselves. Or people with anxiety disorders who may just freak out internally about looking nervous. Lots of examples.
I have autism (previously diagnosed with anxiety disorder), so I thought I'd fit this one, and indeed I do. I wear hoodies (pref. ski sweaters) practically throughout half the year. They help prevent rain and cold on my head, and allow me to reduce external senses. Ski sweaters allow me to protect my hands from cold and dirt.
I also wear glasses which color with the sun (not like dark sunglasses) but work normal inside. When I go outside, I hardly notice the difference with regards to light. I'm not sure what the effect is on blue light filters etc.
I wasn't aware of a racial bias regarding hoodies (I'm what Americans call Caucasian white, but European).
Of course, they didn't go on to generate "deceptive walking" in the wild. They generate video or pictures as deceptive or non-deceptive and they are able to distinguish the generated stuff without labels once they'd trained a network with labeled data.
This might (or might not) serve to prove that these deceptive gestures do exist. The problem of even this is are that there may be a zillion cue characteristic of the subjects being given their directions.
But even if such gestures exist, it seems pretty given that this would be overwhelmed the complexity of images captured in the wild, where many other variable govern behavior.
I grew up on boats and docks. I noticed as a kid that dogs on boats could differentiate people walking on a dock without seeing them. The dogs would become alert and start barking when they heard someone who wasn't normally there. I still wonder if they detect known gaits or the gait of someone who is not walking with a purpose.
I agree, but I was struck by how a dog and I would detect a person at the same time, hearing them walk along the dock. The dog could tell if the person should be there. I could not. Because the dog perked up at the same time as I heard the footsteps, I assumed it was the sound that the dog was using.
The other thing with dogs is that they are empathic and attuned to you. Sometimes they are reacting to what you are communicating to them, before you are even conscious of it.
My cat is able to differentiate between people traversing the stairs to my front door (purely by the sound). The pacing and weight is usually very unique from one person to the next.
She will consistently jump up from my lap to greet a beloved family member at the door, but never gets up for guests.
Congratulations, your AI model can differentiate people intentionally play-acting "deceptive" from people intentionally play-acting "non-deceptive"
After a few years of convtroversry about the low repeatability low-N psychology research published peer-reviewed journals, now we're sharing and upboting non-peer-reviewed non-published psychology papers?
While I could point out several questionable aspects of their methodology, I'm even more troubled by the complete absence of the word 'ethics' from this paper.
Do the people working on these projects never reflect on the ridiculously broad scope of the capabilities they're claiming, the multitude of possibilities for false or inappropriate positives (eg people who are trying to conceal their emotional states, but for wholly valid reasons), the abuses that will be committed by over reliance on such technology and rationalized with spurious claims of objectivity, and the overall rush to make public spaces and the world at large into a giant prison camp where techno-utilitarians pass judgement on the rest of the human race?
Almost every paragraph I look at in this paper makes me roll my eyes. A small sample:
We recruited 88 participants (51 female, 30 male, 7 preferred not to say, age = 20.35) from a university campus.
Wow, fantastically diverse sample.
In the deceptive condition, the participant is supposed to conceal their activities from these “onlookers.” The experimenter stresses that [an experimental prop] is a secret, and because of the amount of money involved, it is really important that the participant keeps anyone from noticing as he/she does these things [with the prop]. In the natural condition, the confederate is wearing a hoodie with the hood pulled down. However, in the deceptive condition, the confederate is wearing the same hoodie with the hood pulled up while wearing sunglasses. Because of this setting, the participant makes a deliberate attempt to conceal his/her activities, resulting in deceptive behavior
We use these findings and consider the following set of gestures:{Hands In Pockets, Looking Around, Touching Face, Touching Shirt/Jackets, Touching Hair, Hands Folded, Looking at Phone}. We chose this set because it includes all the hand gestures observed in the walking videos of participants, and these gestures have been reported to be related to deception
Figure 7. Gesture Features: We present the percentage of videos in which the gestures were observed for both deceptive and natural walks. We observe that deceivers put their hands in their pockets and look around more than participants in the natural condition.
So you got a recruited a bunch of college students and then told some of them to be super-covert while dealing with someone dressed up to look like a stereotype of a shady person, but somehow you imagine this will give an 'honestly deceptive' result that accurately reflects the behavior of deceptive people, rather than the awkward-embarrassed parody of deceptive behavior from someone who feels they're playing a silly game.
Besides all the ethical pitfalls, their research is bad. It's like trying to predict crime based on the observed behavior of people in a school drama class putting on a play about cops and robbers. This is why we have so much of a problem with junk forensic science.
Made me think of the whole labrador dogs cannot hide guilt thing, which is hugely anthropomorphic but it feels like it should be true: if they learned how to model positive and negative human emotion to communicate with us for their advantage, they probably learned shame as well.
Reddit is awash with videos of Labby dogs doing a really bad job of saying "I didn't do it" or of dobbing in the innocent puppy, or whatever.
"Faster. Slower. How We Walk Depends on Who We Walk With, and Where We Live — Men tend to walk differently with other men than with women. And Americans walk faster with children, whereas Ugandans move more leisurely."
Article is a summary of this paper ("Children are not like other loads: a cross-cultural perspective on the influence of burdens and companionship on human walking"):
Figure 1. illustrates the results of an LSTM that's sexually attracted to tall muscular guys, accusing everyone else of being deceptive.
> Deceivers put their hands in their pockets and look around more than the participants in the natural condition.
Doesn't it depend on _context_? If I recall correctly, one of the main giveaways that lead to the identification of the Boston bombers was precisely their composure during the whole situation, which investigators deemed "off" in that context and made them persons of interest.
> Each participant was randomly assigned to walk either naturally or in a deceptive manner, as described below.
So it's like training a "terrorist detection" system by asking your buddies to act "terroristy"? The experiments trains on people displaying behaviors they think convey deception, while deception really tries to convey deceitful behavior.
A better experiment would have been not telling the subjects about the nature of the experiment. Then create a situation where there's a real incentive for the subject to be deceitful. An example would be an experimenter explaining there's a payout for the experiment. The payout must be either monetary or something related. Like saying the study will be at some exotic location, all accomodations taken care of, for a month, and you can bring your significant other.
Then stopping on one answer in particular that disqualifies the subject from the pool of people eligible to that payout.
And then explaining that the experimenter will let this slide but their colleague will probably not; they then instruct the subject and coach them on what they need to avoid saying on several points and before they're done explaining, the "colleague" enters the room and starts going through the checklist.
Now you have incentive to be deceitful and a slightly less crappy experiment than asking people to be deceitful.
Somebody needs to burn down the soft "sciences" and start over. As the other commentors pointed out this a mockery of empiricism and will do real harm when it gets parroted by law enforcement world-wide.
65 comments
[ 3.2 ms ] story [ 110 ms ] threadAlso: what does "we recorded many participants performing tasks involving deceptive walking" mean? Does this AI detect those who perform deceptive walking, or people who are actually deceptive walking?
So yeah! No surprise! They told the participant to act sketchy, and they did! What did the participant do? They put their hands in their pocket and looked around a bunch more than normally. Why? Because they were basically told to! That's not "deceptive walking".
Where is the control for picking up the papers off the floor, or the money off the table? How do we know it's not just that people will walk differently when they are about to crouch down? Maybe that would be an interesting paper by itself, but instead the authors think that this type of research is capable of identifying "deception" with 93% accuracy? Absurd!
There is so much wrong with the experiment, that you could write a rebuttal of almost every decision they made.
This is not deceptive walking, this is honest walking. They are honestly trying to broadcast the signal that they were told to broadcast.
The "deceptive" part is just a frosting layer of prejudice they laid on top.
IQ is only useful in determining pathologies, at the lower end, where any test is usually helpful. For anything else, it's almost entirely snake oil.
If you want to know whether someone is good at something, the best way to test that is to do a test for that particular thing. A test for general inteligence, based on the assumption that general intelligence aptitude is normally distributed in the population, is just not good for anything, and is unlikely to be a reflection of reality, it's pure scientism (instead of science).
OP is treading the same water as
https://arxiv.org/pdf/1611.04135v1.pdf
"Automated Inference on Criminality using Face Images"
Yikes. Anyone getting a feeling of some racial bias in this experimental setup?
https://www.dailytarheel.com/article/2018/04/diversity-0426
>As of fall 2017, UNC-Chapel Hill had a 7.8 percent enrollment rate for Black students, one of the lowest in the state, while just over 22 percent of residents in North Carolina are Black.
Not to mention, there doesn't seem to have been any control in this experiment. What a bunch of nonsense. These students were told to act sketchy, and predictably probably hammed it up to please the experimenters (see: they stuck their hands in their pockets, looked around a lot, etc).
If only 5% of applications to UNC-CH are black, then having 7.8% enrollment is actually good. If 22% of applicants are black, then that is not very good.
You want to increase outreach to underrepresented groups and adjust for biases but then once you have your applicant pool you probably want your selections to mirror the diversity of the applicant pool rather than the state as a whole.
I am scared that our "AI" (really ML) future will be ML trained from bad data and it's pre-crime-fun for anyone that just acts a bit weird or tries to not stand out too much.
I hope there are enough people like you in the institutions where it matters to be all "hey, wait a fuckin SECOND here".
https://www.wired.com/story/best-algorithms-struggle-recogni...
I guess apparently it's still a problem.
I also wear glasses which color with the sun (not like dark sunglasses) but work normal inside. When I go outside, I hardly notice the difference with regards to light. I'm not sure what the effect is on blue light filters etc.
I wasn't aware of a racial bias regarding hoodies (I'm what Americans call Caucasian white, but European).
This might (or might not) serve to prove that these deceptive gestures do exist. The problem of even this is are that there may be a zillion cue characteristic of the subjects being given their directions.
But even if such gestures exist, it seems pretty given that this would be overwhelmed the complexity of images captured in the wild, where many other variable govern behavior.
She will consistently jump up from my lap to greet a beloved family member at the door, but never gets up for guests.
After a few years of convtroversry about the low repeatability low-N psychology research published peer-reviewed journals, now we're sharing and upboting non-peer-reviewed non-published psychology papers?
Do the people working on these projects never reflect on the ridiculously broad scope of the capabilities they're claiming, the multitude of possibilities for false or inappropriate positives (eg people who are trying to conceal their emotional states, but for wholly valid reasons), the abuses that will be committed by over reliance on such technology and rationalized with spurious claims of objectivity, and the overall rush to make public spaces and the world at large into a giant prison camp where techno-utilitarians pass judgement on the rest of the human race?
Almost every paragraph I look at in this paper makes me roll my eyes. A small sample:
We recruited 88 participants (51 female, 30 male, 7 preferred not to say, age = 20.35) from a university campus.
Wow, fantastically diverse sample.
In the deceptive condition, the participant is supposed to conceal their activities from these “onlookers.” The experimenter stresses that [an experimental prop] is a secret, and because of the amount of money involved, it is really important that the participant keeps anyone from noticing as he/she does these things [with the prop]. In the natural condition, the confederate is wearing a hoodie with the hood pulled down. However, in the deceptive condition, the confederate is wearing the same hoodie with the hood pulled up while wearing sunglasses. Because of this setting, the participant makes a deliberate attempt to conceal his/her activities, resulting in deceptive behavior
We use these findings and consider the following set of gestures:{Hands In Pockets, Looking Around, Touching Face, Touching Shirt/Jackets, Touching Hair, Hands Folded, Looking at Phone}. We chose this set because it includes all the hand gestures observed in the walking videos of participants, and these gestures have been reported to be related to deception
Figure 7. Gesture Features: We present the percentage of videos in which the gestures were observed for both deceptive and natural walks. We observe that deceivers put their hands in their pockets and look around more than participants in the natural condition.
So you got a recruited a bunch of college students and then told some of them to be super-covert while dealing with someone dressed up to look like a stereotype of a shady person, but somehow you imagine this will give an 'honestly deceptive' result that accurately reflects the behavior of deceptive people, rather than the awkward-embarrassed parody of deceptive behavior from someone who feels they're playing a silly game.
Besides all the ethical pitfalls, their research is bad. It's like trying to predict crime based on the observed behavior of people in a school drama class putting on a play about cops and robbers. This is why we have so much of a problem with junk forensic science.
Reddit is awash with videos of Labby dogs doing a really bad job of saying "I didn't do it" or of dobbing in the innocent puppy, or whatever.
https://www.nytimes.com/2018/09/19/well/move/faster-slower-h...
Article is a summary of this paper ("Children are not like other loads: a cross-cultural perspective on the influence of burdens and companionship on human walking"):
https://peerj.com/articles/5547/
> Deceivers put their hands in their pockets and look around more than the participants in the natural condition.
Doesn't it depend on _context_? If I recall correctly, one of the main giveaways that lead to the identification of the Boston bombers was precisely their composure during the whole situation, which investigators deemed "off" in that context and made them persons of interest.
> Each participant was randomly assigned to walk either naturally or in a deceptive manner, as described below.
So it's like training a "terrorist detection" system by asking your buddies to act "terroristy"? The experiments trains on people displaying behaviors they think convey deception, while deception really tries to convey deceitful behavior.
A better experiment would have been not telling the subjects about the nature of the experiment. Then create a situation where there's a real incentive for the subject to be deceitful. An example would be an experimenter explaining there's a payout for the experiment. The payout must be either monetary or something related. Like saying the study will be at some exotic location, all accomodations taken care of, for a month, and you can bring your significant other.
Then stopping on one answer in particular that disqualifies the subject from the pool of people eligible to that payout.
And then explaining that the experimenter will let this slide but their colleague will probably not; they then instruct the subject and coach them on what they need to avoid saying on several points and before they're done explaining, the "colleague" enters the room and starts going through the checklist.
Now you have incentive to be deceitful and a slightly less crappy experiment than asking people to be deceitful.