A NSFW Content Recognition Model
Last week our portfolio company Clarifai released a NSFW adult content recognition model.
If you run a web or mobile service that allows users to upload images and videos and are struggling with how to police for NSFW content, you should check it out.
Ryan Compton, the data scientist at Clarifai who built this NSFW model, blogged about the problem of nudity detection to illustrate how training modern convolutional neural networks (convnets) differs from research done in the past.
We are excited about the possibilities that modern neural networks open up for entrepreneurs, developers, and scientists. Our investment in Clarifai is based on our belief that AI/machine learning/neural networks/etc have reached the point of mainstream adoption and usability. And we are seeing more and more use cases for this technology every day.
Solving the problems of content moderators and trust and safety teams at scale, as we discussed here at AVC this past weekend, seems like a particularly good use of this technology.
Comments (Archived):
Great to see you have skin in the game.
should call it #badASSaccurate & appropriate
Are you trying to be puntastic? …’cause that was a good one.
You’re quick on the draw.
Perhaps it should be also known as failing the Stewart test.https://en.wikipedia.org/wi…
Thought about that. If you could train for the emoton of arousal you wouldn’t need to bother with image recognition.
Interesting. Curious to learn how Disqus does this detection already.
They need to make one to detect all the very funny comments and point to them…. AI with a sense of humour!
Piece of cake, use a convoluted reverse hirshfieldian neural network.
There was one one made by a company called Mother Nature. It’s always active on AVC 🙂
they are not going to say, that would be bad….
What I love about AVC: Fred always keeps me abreast of the situation.
From Clarifai’s blog, it’s a great read.* http://blog.clarifai.com/wh…
Looking forward to clarifai’s pun recognition model.
Ouch!
The beast of Natural Language understanding in AI is a different one from the image and verbal speech recognition problem that Deep Learning convolutional and recursive neural networks are well on their ways to solving.
NSFW = Not Safe For Workwhat are we saying here?S = Safe, as meaning (Not) ‘appropriate’?and because…W = Work, as in a space populated by many (with differing ‘values’)?
My wife pointed out the other day how much more casual nudity we see on TV – as a result of going 100% on-demand streaming – Netflix/Hulu/Amazon.The flip side is we see almost no violence in our choices – no hookers being strangled, no shocking gun deaths, no gratuitous slashing… a staple of prime time network TV.She thinks this is a good thing – and commented on how many brutal deaths our kids must have seen on TV growing up – even with our “good” parenting. I’d much rather have a bit of nudity slip through than having kids grow up with a constant stream of gratuitous violence.
So true. American TV is laced with violence – but your career dies a silent death if you dare show a boob.
“boob”? You mean female breast? As men, we should be protective and show caring and respect.
shouldn’t everyone do that regardless of gender/sex
About the most men can hope for is that, before the shooting starts, both sides let all the women and children get out of the way.
In other news Intel cuts 11% of workforce. 12,000 people. WTHA bad earnings call and the CEO isn’t fired but 12,000 workers lose their jobs?http://www.forbes.com/sites…
Great innovation. Will be interesting to see if the spammers et al try to find a work around.
basically, they will have to work along the lines of email spammers after bayesian filters, and unlike bayesian filters, every time something gets through and is tagged/untagged, the system develops much better immunity rather than wholesale probability shifts.In essence a spammer will have to send one million images out in order to guarantee a .0001% success rate. (that would be 1 image posted and sticking) over time they would need to build their own convolutional type networks to also predict what kind of images would get through as well. Both propositions are expensive, so probably spamming will consolidate into groups who can afford to do so.
The pricing model is interesting. It’s cheaper to get multiple 20k per month accounts and 1 40k per month account (to get the prioritized support) than it is to get the 250k per month account at a certain volume level…
Good! I’ll keep the post AND its links.One link has a lot of examples of using the Web utility cURL that permits scripting interactions by a client between the client and a Web server. GOOD cURL examples!Yesterday I spent a good chunk of time writing scripts to use cURL and finally got the scripts to work on all but one of the Web sites I tried.But in getting cURL to work, there was mud wrestling involved, and some frustration at the inscrutable syntax, semantics, and actual errors of some of the cURL documentation. If the documentation writers want to use Bachus-Naur form (BNF) to specify syntax, fine, but in that case they should learn actually to use BNF.. Else it’s not just better but essential that they give good working examples. Well, the Clarifai page has some such examples!One of the errors in the cURL documentation was the statement that, in the command line arguments for cURL, for a string that contains blanks, one should enclose the string in “single quotes.” WRONG! Have to use double quotes.So, I did the usual for such mud wrestling: Tried to do something really simple that didn’t have many cURL options where had to resolve documentation ambiguities. So, for just the few options, just try all the possibilities of the ambiguities one at a time until the thing works. Then add more options one at a time, by such enumeration resolve all their ambiguities, and repeat. Got it to work. Mud wrestling — use the TIFO method, try it and find out. So f’get about being at all rational, regard the documentation as no more than just vague hints, and just TIFO-it!So far, given a image, Clarifai reports back a number from 0 to 1, that is, 0% to 100%, with less than 15% meaning likely porn and greater than 85% meaning safe for work. That’s 70% of ambiguity in the middle!Clarifai says that they have a detector, filter, whatever, that recognized wedding pictures. Okay. But somewhere on the Internet is a wedding picture of the bride and her bridesmaids where they are all standing there, posing for the camera, really happy, smiling, and have their gowns pulled up to their waists illustrating that all they were wearing was their gowns and, presumably, were all ready for the wedding night! So, it’s wedding picture porn! So, on that picture, the poor AI-bot will keep saying, “It’s a wedding picture!”; “No, it’s porn!”; “No, it’s a wedding picture!”, … until its circuits overheat and it throws smoke and sparks like some Star Wars R2-D2 hit with a blaster and then falls over “Clunk”!So, we’re back to the challenge of getting meaning from Internet content, in this case, still images. E.g., for porn, we remember the old statement of some politician, “I know it when I see it.” — and you are saying that you’ve seen a lot of it, right?The recent computer science (time out to do an up-chuck — science used to be respectable) machine learning techniques are one approach.But there are some quite different approaches based on some more general and powerful math (yes, that I am using — don’t look for it in the libraries; I derived it; next to no computer science professors have the prerequisites for it; since it’s theorems and proofs, I know that at least mathematically it’s correct). Ah, I prefer what I have!Uh, are they recognizing porn or just, say, Jenna Jameson? Maybe they are just recognizing the partially shut eyes of submissiveness of sexual arousal! Maybe they are just matching against a collection of images from 4chan! Maybe they are just doing character recognition looking for “MetArt”?Porn is nearly always characterized by just views of just a few square inches of a human body, usually female. So, given a whole picture, just partition it into sub pictures and do shape recognition on those individually. Then the whole picture is porn if and only if (iff — thanks to P. Halmos) some such sub picture is. That stands to be more accurate.
Had fun throwing some pictures at Clarifai’s NSFW sample page: Blake Lively, a naked sphynx cat, a naked couple and a Tunick naked group. The naked couple was the only one tagged NSFW. Amazing.Strolling naked through the fields is safe for work, Game of Thrones should be OK..
Clarifai is impressive. I tested it with some unusual images like hand drawn sketches and diagrams. The tag results are surprisingly accurate, particularly since I’ve tried the same exercise with Google’s image recognition and had no results. Excellent work.
Keep your dongle away from me
the digital battle will play out if the NSFW entities think the blocked entities are a target market they can make money in. I would not invest in a technology that enabled the pornographers. I remember when Mr. Skin was started up-and I didn’t invest then and don’t regret it even though the business has been highly successful. (Was started by a clerk on the CME floor)
Lol today’s topic is a field day for you! 🙂
I would not invest in a technology that enabled the pornographers.Is that because you think it makes you unseemly as an investor? Or couldn’t face your family or peers? Or because you think there is something particularly wrong with pornography?There is this association with pornography as being unseemly, wrong and bad, even though it’s quite obvious that the large majority of men either view or have viewed considerable pornography. And that in some respects it’s quite healthy.Considering the other actual illegal vices that people have (pot – not legal in many places, alcohol and drug abuse (prescription or otherwise), gambling, guns!, that are often celebrated and tolerated and have actual harm I find this attitude ironic but totally expected. [1][1] Not that I would feel comfortable being involved myself of course…
Meh, just don’t want to make money that way. Has nothing to do with anyone else. I don’t judge people that do unless they are using slaves or something like that. Sex industry is full of that stuff and I don’t want to be a part of it.
excellent market analysis and product execution.
Nice work. Understanding and implementing cultural diversity is key.
Yep after I wrote that I remembered that you are on some important boards (public and private) and a well respected member of your community. And nothing would kill your standing (which you derive enjoyment and benefit from and I say that as a positive not as a dig by the way) than being associated with the wrong stuff.That said at my bar mitzvah back in the day the biggest gift that I got was from a neighbor who owned porn shops. $1000 in today’s dollars. He also gave a great deal of money to the synagogue (kind of like the mafia giving money to churches). I realized then that houses of worship couldn’t exist and serve the poor without people and somewhat dirty money keeping them floating. They needed all sorts of members. In other words the guy could cleanse himself by not attending synagogue but giving it money. The poor person who paid money but showed up everyday was viewed a bit differently. Made total sense to me.
the boards I am on and all that stuff have nothing to do with it. Just don’t feel comfortable investing in it. Zen/Karma thing probably. maybe it’s having two daughters. Believe me, I am no angel.You must have had a helluva bar mitzvah. ha.
I actually turned down some business in that industry as well. The real issue is that while the high end seems harmless, it goes downhill really quickly.
In my dad’s group of friends it was not uncommon to give the gift after seeing the affair that was made (write the check at the affair). They are a tough bunch of immigrants. I hated it hated hated hated being the center of attention. There was a valium standing by the night before in case I needed it (I didn’t). I hated everything about it. But I knew enough without being told to press the flesh with all people attending to make sure they gave big gifts. I rose to the occasion. Back in that day it was a band (no dj’s never ever done) and the film was color super 8 without sound. The special effects were so primitive compared to todays.Leading up to the affair I remember my dad saying “eh, if the cook has a fight with his wife you’ll have a bad affair”. Didn’t understand that until I got older and in relationships.