This is very exciting as I toyed with this problem around 2000-2001. I speculated it should be easy to recognize porn because the images had recurring patterns. I made a list of a bunch of them that should at least reduce a moderator's workload. Not going into details but the best foresight I had was recognizing the Got Milk commercials would throw the classifications off. It did lol.
In any case, we just had things like fuzzy logic, SVM's, primitive ANN's, and so on. I could describe and illustrate the patterns with ease but couldn't encode them into an algorithm for the life of me. That the modern tech created in the image set some of the same exact patterns my brain spotted is pretty amazing. The researchers also have a ton of training data to better illustrate NSFW and SFW. That they're throwing in lots of SFW is very smart as it's easier to bias this topic than most.
All I can say I keep up the good work. Been neat to see an insurmountable-on-Pentium2 problem get stomped by the CNN's. I still want to see them applied more to hunting source code defects, host forensics, and intrusions.
In any case, we just had things like fuzzy logic, SVM's, primitive ANN's, and so on. I could describe and illustrate the patterns with ease but couldn't encode them into an algorithm for the life of me. That the modern tech created in the image set some of the same exact patterns my brain spotted is pretty amazing. The researchers also have a ton of training data to better illustrate NSFW and SFW. That they're throwing in lots of SFW is very smart as it's easier to bias this topic than most.
All I can say I keep up the good work. Been neat to see an insurmountable-on-Pentium2 problem get stomped by the CNN's. I still want to see them applied more to hunting source code defects, host forensics, and intrusions.