3 ms·

Translation for those unfamiliar: they used machine learning. And while SVM is not "advanced enough" compared to "deep learning"[0], it's a very powerfull class

Translation for those unfamiliar: they used machine learning. And while SVM is not "advanced enough" compared to "deep learning"[0], it's a very powerfull classificator/estimator on its own. To OP: maybe today with a huge GPU, lots of data and more complex models you could get something going. Do you think a more advanced model could really pull this through? Maybe using other features -- perhaps coupling with a thermal camera can help detect changes in temperature that could signal the person is lying. [0] quotes because I've heard this complaint a few times already, and it's usually from people that have bought the buzzword but usually don't understand much about machine learning in general.

misiti3780 8y ago
I havn’t been following this story really closely today, but here is my two sense (albeit outdated): In the research I was working on, there were two types of deception/lies: high stakes (aka suicide bomber or shoe bomber) and low states (you have a pound of cocaine strapped to your leg). If you want to build a lie detector on facial data, you need consider things like 1. It is important to have a robust data set that contains lots of example of both for training and cross-validation 2. False positives can be very expensive (because you start pulling people aside, lines build up, people get pissed, etc) The project I worked on was trying to solve a specific problem: replace Behavior detection officers [1]: These are people that sit inconspicuously at the TSA checkpoints and decide who gets to get pulled aside for further screening and who goes through (I do know anything else about what they do and where they are, numbers etc.) Their training is advanced, laborious, and expensive for TSA. They are, of course, trying to catch terrorists (aka the high stakes lies) but in reality they are dealing with the low stakes lies. The only terrorist I can think of since 9/11 that got close to taking down a plane was Richard Reid, and they missed him so they seem to be 0/1 in high states lies). Anyways, the problem with building their replacement is we do not have a lot of good examples of high stakes lies in video data. There are some (Scott Peterson, etc) but not enough. Also, the models they were built off of were "micro-expression" [2] detectors, and micro-expressions show up differently across cultures (maybe the consensus has changed since I was modeling this stuff) . So, in short, the project I was on failed because the data sucked and even if we did build a model, it was overfit to the data we trained it on. I am not sure how deep learning and more advanced GPUs would solve that. Maybe they found better data, but I would be skeptical. I honestly would be surprised if they actually did much better than random (ROC > .5) Sorry for the rambling - end of the day for me on the east coast. [1] https://www.tsa.gov/news/testimony/2013/11/14/tsa-behavior-detection-and-analysis-program https://www.tsa.gov/news/testimony/2013/11/14/tsa-behavior-d... [2] https://en.wikipedia.org/wiki/Microexpression https://en.wikipedia.org/wiki/Microexpression
- dr_zoidberg 8y ago
  Thanks for the answer! I was thinking more along the lines of data augmentation. All in all, I don't think microexpressions alone are a good feature, from your description it's clear that they're problematic on a number of fronts. The article also mentions microexpressions as the basis for this new system, and that they worked on a slow universe of 30 people, with ~76% accuracy. What remains to be seen is if the new, larger deployment will work better. Also, they may be working more focused on the low stakes lies rather than the high stakes. What I was thinking that "more compute power" could bring to the table would be deeper and more complex models, faster training algorithms[0] and data augmentation. But then again, all that relies on having good data to begin with. [0] I worked with neural nets in the past (~10 years, I pretty much left the field right before the deep learning big bang), and coming back to the field I'm pretty much amazed at how training a 7 layer deep MLP used to take a few minutes on CPU back in 2010, and now can be done in seconds (still on CPU). And while there's a part of microarchitectural improvement to account for, that could at best be a 5x speedup -- the rest is owed to the newer training algorithms. Once again, thanks for your answer!

misiti3780 8y ago

I havn’t been following this story really closely today, but here is my two sense (albeit outdated): In the research I was working on, there were two types of deception/lies: high stakes (aka suicide bomber or shoe bomber) and low states (you have a pound of cocaine strapped to your leg). If you want to build a lie detector on facial data, you need consider things like 1. It is important to have a robust data set that contains lots of example of both for training and cross-validation 2. False positives can be very expensive (because you start pulling people aside, lines build up, people get pissed, etc) The project I worked on was trying to solve a specific problem: replace Behavior detection officers [1]: These are people that sit inconspicuously at the TSA checkpoints and decide who gets to get pulled aside for further screening and who goes through (I do know anything else about what they do and where they are, numbers etc.) Their training is advanced, laborious, and expensive for TSA. They are, of course, trying to catch terrorists (aka the high stakes lies) but in reality they are dealing with the low stakes lies. The only terrorist I can think of since 9/11 that got close to taking down a plane was Richard Reid, and they missed him so they seem to be 0/1 in high states lies). Anyways, the problem with building their replacement is we do not have a lot of good examples of high stakes lies in video data. There are some (Scott Peterson, etc) but not enough. Also, the models they were built off of were "micro-expression" [2] detectors, and micro-expressions show up differently across cultures (maybe the consensus has changed since I was modeling this stuff) . So, in short, the project I was on failed because the data sucked and even if we did build a model, it was overfit to the data we trained it on. I am not sure how deep learning and more advanced GPUs would solve that. Maybe they found better data, but I would be skeptical. I honestly would be surprised if they actually did much better than random (ROC > .5) Sorry for the rambling - end of the day for me on the east coast. [1] https://www.tsa.gov/news/testimony/2013/11/14/tsa-behavior-detection-and-analysis-program https://www.tsa.gov/news/testimony/2013/11/14/tsa-behavior-d... [2] https://en.wikipedia.org/wiki/Microexpression https://en.wikipedia.org/wiki/Microexpression

dr_zoidberg 8y ago
Thanks for the answer! I was thinking more along the lines of data augmentation. All in all, I don't think microexpressions alone are a good feature, from your description it's clear that they're problematic on a number of fronts. The article also mentions microexpressions as the basis for this new system, and that they worked on a slow universe of 30 people, with ~76% accuracy. What remains to be seen is if the new, larger deployment will work better. Also, they may be working more focused on the low stakes lies rather than the high stakes. What I was thinking that "more compute power" could bring to the table would be deeper and more complex models, faster training algorithms[0] and data augmentation. But then again, all that relies on having good data to begin with. [0] I worked with neural nets in the past (~10 years, I pretty much left the field right before the deep learning big bang), and coming back to the field I'm pretty much amazed at how training a 7 layer deep MLP used to take a few minutes on CPU back in 2010, and now can be done in seconds (still on CPU). And while there's a part of microarchitectural improvement to account for, that could at best be a 5x speedup -- the rest is owed to the newer training algorithms. Once again, thanks for your answer!