13 ms·

This is absolutely fascinating. It's mesmerizing to see this NSFW detection applied in reverse, and it's even more interesting to observe your mind react to th

by niftich 10y ago

This is absolutely fascinating. It's mesmerizing to see this NSFW detection applied in reverse, and it's even more interesting to observe your mind react to the generated images. You can see the sort-of-mons pubis patterns, the maybe-pubic hair, the perhaps-breasts and the suspiciously phallic appendages, complete with realistic colors. Interestingly, all exposed skin suggests that the training dataset for the NSFW detection was skewed towards caucasians, given how the synthesized images are near-completely devoid of skin tones other than light pink. Perhaps this is a good visual indication of unintentional 'bias' in datasets?

walrus01 10y ago
Who wants to volunteer to acquire and organize a several thousand jpg set of "ebony" category porn and submit it? If training material is required, there's undoubtedly a lot of categorized porn for all possible races in the world already, it's just a matter of getting volunteers to do it.
- niftich 10y ago
  I'm actually really curious whether nudes of dark-skinned people pass the default model at all. Luckily, detector can be seeded to work with a different model, so we can have several of them; but this would be a useful pull-request.
- devonkim 10y ago
  It may be more beneficial to train for just body feature recognition instead of skin tone because last I saw, from the neck down women have only so much variance across body shapes. We already have a huge amount of homemade porn by people all across the world, but the dataset is not very well labeled compared to professional mainstream porn and that will impact the research. In other posts I've seen, generative networks produced biases unrelated to what people consider NSFW such as brighter lip color, so I think restricting features related to anything above the neck or below the knees from impacting the CNNs would be helpful.
  - otakucode 10y ago
    I would be curious to know what these images would look like if it were trained on a corpus including both the dataset used AND a dataset of hand-drawn pornography, including anime and 'cartoony' porn. I'd also be curious to know how often this sort of network would bar educational content. There is an important difference between nudity and pornography. It's a distinction that I would think would be very difficult to train any automated system on, but I am prepared to be surprised.
gxs 10y ago
You just had to throw race in there. It's the easiest cheapest way nowadays to sound insightful. Edit: I think I came across harsher than I meant to. All I meant was that we could have simply said I wonder how it deals with different skin tones. Instead, we get into race and biases etc. Interesting that regarding universities being sued for lack of cc in their free online content everyone was up in arms, but mention the word race and everyone goes on high alert.
- hasbroslasher 10y ago
  Commenting about how conversations of race aren't welcome in a given discussion is the easiest way to sound like you're missing the point
- niftich 10y ago
  I couldn't reply in time before you made your edit, but truthfully I was going to post the same thing: I was struggling with how to phrase my original comment, but I was looking for a shorthand way of saying "people with light pink and white skin" without coming off as either callous or overly euphemistic. Really, my curiosity stems from being new to how neural networks work -- naively, I've grown accustomed to processes where the outputs reflect the input, and at first blush to me, supposing a varied dataset, this didn't appear to be the case. Those who replied to me provided persuasive arguments as to why the hues of the synthesized images may not necessarily imply that the training dataset wasn't that varied after all.
  - gxs 10y ago
    It's cool man I'm just burned out on the subject and took it out unfairly on your comment. It's a valid point you made, it was just the wording that pushed a button.
- eric_h 10y ago
  I upvoted you because of your edit, but I thought I'd add that "race" is a completely social construct and has no basis in science. Skin tone would indeed have been a more appropriate term to use. [edit, removed a parenthetical as it undermined my point]
gabrielgoh 10y ago
glad you like it :) (author here) when running the experiments, i noticed the neural net responds much more strongly to contrast than color. this is consistent with previous work where the the first layers of a conv net isolate colors in a few filters. i wouldn't read too much into this
- alimw 10y ago
  This is amazing! I'm particularly pleased to see how it comes out because I remember wanting to do precisely this (generative model plus porn) back in 2000. I had recently seen Geoff Hinton give a talk, and back then I had half my focus on maths and computers and half on art and design. Of course I never got close to doing anything about it :) Maybe it wouldn't have been possible.
- otakucode 10y ago
  I'm not terribly surprised - humans do too. Contrast is well established as being far more important for perceived image 'quality' than color fidelity.
aab0 10y ago
It could also reflect what is most distinguishable. Which is easier for a NN to confidently distinguish: black pubic hair on black skin, or black pubic hair on white skin? Darker nipples on black skin, or darker nipples on white skin? etc You're doing gradient ascent on confidence of classification, not simply trying to find a plausible input, but the maximal input. There's no reason to expect this to be racially unbiased as there are simple objective reasons that higher contrast would be useful. (Similarly, I would not be surprised if a face recognition NN worked better on Europeans rather than Chinese, for the prima facie reason that they have more variable facial features and other aspects like multiple hair colors other than black.) And since Yahoo needs to detect porn of all races and there's plenty of black porn out there, it would be odd if their porn detector had such a huge gaping hole in it.
- slackstation 10y ago
  > it would be odd if their porn detector had such a huge gaping hole in it phrasing! I couldn't resist.
  - aab0 10y ago
    That was deliberate. I couldn't resist either.
  - FuckOffNeemo 10y ago
    Have some mindless up votes Sterling.
- HSO 10y ago
  > Similarly, I would not be surprised if a face recognition NN worked better on Europeans rather than Chinese, for the prima facie reason that they have more variable facial features and other aspects like multiple hair colors other than black Hair color, yes, but facial feature variability? I think your perception of more variability in European faces than Chinese ones says more about where/how you grew up than about the actual variability. To someone who grew up in an Asian monoculture, without exposure to media with European faces, all Europeans would look alike too.
pbhjpbhj 10y ago
Re your last point: Strong pink is probably the most common colour across all people for the most intimate parts of the genitals as AFAIK the inner folds and spread appearance of all types of vulva are pink. All tongues and palms are pink too [AFAIK], they probably feature highly in porn compared to in other images. If you were then looking at pictures of black women, say, and wanted to find the porn images then looking for certain pinks in those images would be a good first pass I feel [you'd probably get lots of false-positives for mouth pictures]. Even those with brown and black skin can flush pink when aroused too, that's got to feed in to the results. [Do men have pigmented glans penis? Not something I've ever really needed to know. [2]] FWIW https://open_nsfw.gitlab.io/nsfw/prob_0001_1000_10.0_0.04__144.jpg https://open_nsfw.gitlab.io/nsfw/prob_0001_1000_10.0_0.04__1... doesn't look especially "caucasian" (ie white European) to me. There seems to be a good mix of pretty coffee-y colours to me that I'd associate with Asia. >Perhaps this is a good visual indication of unintentional 'bias' in datasets? // So, maybe it's a colour-of-genitals bias rather than a dataset bias? [1] http://www.iflscience.com/health-and-medicine/why-are-your-nipples-and-genitals-different-color-rest-your-skin-0/ http://www.iflscience.com/health-and-medicine/why-are-your-n... [2] http://www.interestingarticles.com/mens-health-issues/penis-skin-of-a-different-color-whats-normal-and-whats-not-5948.html http://www.interestingarticles.com/mens-health-issues/penis-...
- vidarh 10y ago
  > All tongues and palms are pink too That's not true for palms.
inlined 10y ago
I also think there was a clearer bias towards mushroom shaped objects than O'Keef work. It made me wonder if nude men are considered more pornographic than men.
- dspillett 10y ago
  > It made me wonder if nude men are considered more pornographic than men. In some places, definitely yes. The laws and regulations imposed by different countries, states, media companies, and other bodies, and how they are sometimes "over interpreted" when being careful not to trip over the rules, do vary but there is some bias towards not showing the penis at least. In some contexts genitalia is banned but breasts and buttocks are fine leaving more room for a sexual-ish image of woman to be acceptably published. In other contexts genitalia can be show as long as it isn't in a obviously aroused state, but other shows of arousal (blushing, other body language, suggestive word choice, ...) are fine, which will likely bias against male genitalia being publishable unless close up (and in contexts where the question is even raised a close-up shot is less likely). (Massive generalism here, but...) there is also a general opinion that the use of the female form can be defended as artistic more than the male - though I couldn't tell you if this is a symptom of other cultural biases or a cause of some of them. Attitudes to sexual preference are going to factor in too: those anti-homosexual opinions/feelings (conscious or otherwise) are more likely to react strongly to images of male sexuality than female (the old joke "I support homosexuality as long as both the chicks are hot" has some basis in reality). This is more common in men in my experience, especially man of certain age ranges and backgrounds which are currently over represented in parts of the "traditional" media industry and other seats of influence, making the bias more visible. All that and a number of other interrelated factors I don't have time to type out make it less common to see a full-body or otherwise genitalia-containing shot of a naked male than one of a naked female - so when one is shown it is more likely to illicit a stronger reaction. (I avoided using examples of differences in the consumption of pornographic output between different sexualities, as there is a lot of reporting bias in studies in that area both in terms of what questions the studies ask, how truthful the subjects answer, and what gets selected for publication)

by niftich 10y ago

walrus01 10y ago

Who wants to volunteer to acquire and organize a several thousand jpg set of "ebony" category porn and submit it? If training material is required, there's undoubtedly a lot of categorized porn for all possible races in the world already, it's just a matter of getting volunteers to do it.

niftich 10y ago
I'm actually really curious whether nudes of dark-skinned people pass the default model at all. Luckily, detector can be seeded to work with a different model, so we can have several of them; but this would be a useful pull-request.
devonkim 10y ago
It may be more beneficial to train for just body feature recognition instead of skin tone because last I saw, from the neck down women have only so much variance across body shapes. We already have a huge amount of homemade porn by people all across the world, but the dataset is not very well labeled compared to professional mainstream porn and that will impact the research. In other posts I've seen, generative networks produced biases unrelated to what people consider NSFW such as brighter lip color, so I think restricting features related to anything above the neck or below the knees from impacting the CNNs would be helpful.
- otakucode 10y ago
  I would be curious to know what these images would look like if it were trained on a corpus including both the dataset used AND a dataset of hand-drawn pornography, including anime and 'cartoony' porn. I'd also be curious to know how often this sort of network would bar educational content. There is an important difference between nudity and pornography. It's a distinction that I would think would be very difficult to train any automated system on, but I am prepared to be surprised.

gxs 10y ago

You just had to throw race in there. It's the easiest cheapest way nowadays to sound insightful. Edit: I think I came across harsher than I meant to. All I meant was that we could have simply said I wonder how it deals with different skin tones. Instead, we get into race and biases etc. Interesting that regarding universities being sued for lack of cc in their free online content everyone was up in arms, but mention the word race and everyone goes on high alert.

hasbroslasher 10y ago
Commenting about how conversations of race aren't welcome in a given discussion is the easiest way to sound like you're missing the point
niftich 10y ago
I couldn't reply in time before you made your edit, but truthfully I was going to post the same thing: I was struggling with how to phrase my original comment, but I was looking for a shorthand way of saying "people with light pink and white skin" without coming off as either callous or overly euphemistic. Really, my curiosity stems from being new to how neural networks work -- naively, I've grown accustomed to processes where the outputs reflect the input, and at first blush to me, supposing a varied dataset, this didn't appear to be the case. Those who replied to me provided persuasive arguments as to why the hues of the synthesized images may not necessarily imply that the training dataset wasn't that varied after all.
- gxs 10y ago
  It's cool man I'm just burned out on the subject and took it out unfairly on your comment. It's a valid point you made, it was just the wording that pushed a button.
eric_h 10y ago
I upvoted you because of your edit, but I thought I'd add that "race" is a completely social construct and has no basis in science. Skin tone would indeed have been a more appropriate term to use. [edit, removed a parenthetical as it undermined my point]

gabrielgoh 10y ago

glad you like it :) (author here) when running the experiments, i noticed the neural net responds much more strongly to contrast than color. this is consistent with previous work where the the first layers of a conv net isolate colors in a few filters. i wouldn't read too much into this

alimw 10y ago
This is amazing! I'm particularly pleased to see how it comes out because I remember wanting to do precisely this (generative model plus porn) back in 2000. I had recently seen Geoff Hinton give a talk, and back then I had half my focus on maths and computers and half on art and design. Of course I never got close to doing anything about it :) Maybe it wouldn't have been possible.
otakucode 10y ago
I'm not terribly surprised - humans do too. Contrast is well established as being far more important for perceived image 'quality' than color fidelity.

aab0 10y ago

It could also reflect what is most distinguishable. Which is easier for a NN to confidently distinguish: black pubic hair on black skin, or black pubic hair on white skin? Darker nipples on black skin, or darker nipples on white skin? etc You're doing gradient ascent on confidence of classification, not simply trying to find a plausible input, but the maximal input. There's no reason to expect this to be racially unbiased as there are simple objective reasons that higher contrast would be useful. (Similarly, I would not be surprised if a face recognition NN worked better on Europeans rather than Chinese, for the prima facie reason that they have more variable facial features and other aspects like multiple hair colors other than black.) And since Yahoo needs to detect porn of all races and there's plenty of black porn out there, it would be odd if their porn detector had such a huge gaping hole in it.

slackstation 10y ago
> it would be odd if their porn detector had such a huge gaping hole in it phrasing! I couldn't resist.
- aab0 10y ago
  That was deliberate. I couldn't resist either.
- FuckOffNeemo 10y ago
  Have some mindless up votes Sterling.
HSO 10y ago
> Similarly, I would not be surprised if a face recognition NN worked better on Europeans rather than Chinese, for the prima facie reason that they have more variable facial features and other aspects like multiple hair colors other than black Hair color, yes, but facial feature variability? I think your perception of more variability in European faces than Chinese ones says more about where/how you grew up than about the actual variability. To someone who grew up in an Asian monoculture, without exposure to media with European faces, all Europeans would look alike too.

pbhjpbhj 10y ago

Re your last point: Strong pink is probably the most common colour across all people for the most intimate parts of the genitals as AFAIK the inner folds and spread appearance of all types of vulva are pink. All tongues and palms are pink too [AFAIK], they probably feature highly in porn compared to in other images. If you were then looking at pictures of black women, say, and wanted to find the porn images then looking for certain pinks in those images would be a good first pass I feel [you'd probably get lots of false-positives for mouth pictures]. Even those with brown and black skin can flush pink when aroused too, that's got to feed in to the results. [Do men have pigmented glans penis? Not something I've ever really needed to know. [2]] FWIW https://open_nsfw.gitlab.io/nsfw/prob_0001_1000_10.0_0.04__144.jpg https://open_nsfw.gitlab.io/nsfw/prob_0001_1000_10.0_0.04__1... doesn't look especially "caucasian" (ie white European) to me. There seems to be a good mix of pretty coffee-y colours to me that I'd associate with Asia. >Perhaps this is a good visual indication of unintentional 'bias' in datasets? // So, maybe it's a colour-of-genitals bias rather than a dataset bias? [1] http://www.iflscience.com/health-and-medicine/why-are-your-nipples-and-genitals-different-color-rest-your-skin-0/ http://www.iflscience.com/health-and-medicine/why-are-your-n... [2] http://www.interestingarticles.com/mens-health-issues/penis-skin-of-a-different-color-whats-normal-and-whats-not-5948.html http://www.interestingarticles.com/mens-health-issues/penis-...

vidarh 10y ago
> All tongues and palms are pink too That's not true for palms.

inlined 10y ago

I also think there was a clearer bias towards mushroom shaped objects than O'Keef work. It made me wonder if nude men are considered more pornographic than men.

dspillett 10y ago
> It made me wonder if nude men are considered more pornographic than men. In some places, definitely yes. The laws and regulations imposed by different countries, states, media companies, and other bodies, and how they are sometimes "over interpreted" when being careful not to trip over the rules, do vary but there is some bias towards not showing the penis at least. In some contexts genitalia is banned but breasts and buttocks are fine leaving more room for a sexual-ish image of woman to be acceptably published. In other contexts genitalia can be show as long as it isn't in a obviously aroused state, but other shows of arousal (blushing, other body language, suggestive word choice, ...) are fine, which will likely bias against male genitalia being publishable unless close up (and in contexts where the question is even raised a close-up shot is less likely). (Massive generalism here, but...) there is also a general opinion that the use of the female form can be defended as artistic more than the male - though I couldn't tell you if this is a symptom of other cultural biases or a cause of some of them. Attitudes to sexual preference are going to factor in too: those anti-homosexual opinions/feelings (conscious or otherwise) are more likely to react strongly to images of male sexuality than female (the old joke "I support homosexuality as long as both the chicks are hot" has some basis in reality). This is more common in men in my experience, especially man of certain age ranges and backgrounds which are currently over represented in parts of the "traditional" media industry and other seats of influence, making the bias more visible. All that and a number of other interrelated factors I don't have time to type out make it less common to see a full-body or otherwise genitalia-containing shot of a naked male than one of a naked female - so when one is shown it is more likely to illicit a stronger reaction. (I avoided using examples of differences in the consumption of pornographic output between different sexualities, as there is a lot of reporting bias in studies in that area both in terms of what questions the studies ask, how truthful the subjects answer, and what gets selected for publication)