It's 90% my dude
The J-AI returns
Adam Mitchell
Angel Torres
transfer learning with vgg16 (even tried with inception, resnet and some others for lulz), tried it with multiple subsets of data collected from prevoius pol threads, manual collection, twitter/vk crawlers.
Had little to do with facial features, it was based on selfies without any real preprocessing so angle, lighting, maybe colors, accessories, presentation etc would all impact the outcome
Nicholas Lopez
98% for a CNN. I know this is major ethnicities, but getting even 75-80% is more than enough for the purpose of a funny app used by Zig Forums and teenagers having fun. I currently use images that often contain most of the hair and ears at 224x224px and like you said yourself, I think the problem is in the data, since Zig Forums is pretty good at detecting jews and current NNs achieve human level performance on most facial recognition tasks.
Aaron Williams
read my previous posts again
you will never even reach 55% unless you use some terribly biased dataset consiting of nothing but caricatures at which point the entire thing is redundant anyways
Carson Taylor
J-dar was a lot of fun. If only there was a way to pull its index and mirror it.
Bump. As you said it would be worth it just for the sake of letting other people use it.
If you're so adamant, how about you release your version and let OP use it as an example? If it's as you said, a waste of time, then you could help OP save some time so he can begin his attempt at it. Maybe he will find something you missed. Help him by helping him save some time.
You yourself mentioned caricatures, but even admit an experienced Zig Forums user can detect a kike. The less pronounced features however are still detectable. That enough is a way of helping a machine detect kikes.
Connor Evans
MOST jews aren't difficult to identify using physical features, voice/nasal tone, and behavior, which is the most telling. It would be nice to have an app to help those who need it until they can spot jews without it. I hate seeing people post Lauren Southern or some other jew when I can they are jews. Why can't most people see it? That's what needs to be addressed.
Owen Reed
Why not an app called 'Your Ugly Self' that turns you into a super nasty looking kike? It would be fun for zoomers and a warning at the same time.
Jaxson Walker
You're trying to classify those of jewish descend based solely on their facial features. Of course that means you can also make a distinct difference between those who aren't jewish and those who are. A problem encountered during the classification task of jews is that their features cluster a lot with other ethnicities. Most Jewish that are of Ashkenazi descend have mixed features, some predominantly European / Caucasian. It was hardest to classify those of Caucasian descent since there was the most diversity of facial features in those groups. You could split it up into eastern and western european features and you could still get a very high accuracy of 90%+ (that coupled with other racial groups).
The reason why caricatures are mentioned is that the classifier itself is only very reliable in regards to Jewish caricatures, or the Seraphim Jewish group. It's not really good at all to identify mixed race Jews or Ashkenazi Jews which are the majority of those who have Jewish blood in them. Due to the large overlap in clusters that are formed to classify, the accuracy when it comes to the majority of Jews is quite low.
The reason why caricatures are bad is because you will give the machine learning algorithm a bias to only detect caricatures. We don't want an algorithm that can only classify those who look like caricatures. The reason for that is you put in bias (caricatures) that are an obvious cluster of their own and then say 'look it works on other caricaturist looking jews'. But it wouldn't detect a Jew of Ashkenazi descend who has some European features. So it's rather useless because Jewish people are so mixed in particular with Caucasian people that they are very difficult to detect.
You can indeed as a human see some difference because you notice a faint mix of Jewish / Arabic features and white features. The people who have those features are often Jewish. It's difficult for a machine to see the same thing in a mathematical vectorized way due to the high overlap with other races problem.
Wyatt Perez
My trash, unsorted dataset gave me 64% validation accuracy on average. 50% is just random guessing, which would imply zero difference in general facial patterns. Do you believe that neural networks aren't at a level currently where they can detect two distinct ethnicities? Or do you believe the task of creating a high quality dataset is too difficult?
Couldn't this be solved through a more granular dataset? Wouldn't the main thing blocking the networks generalization be the large variance between the Jews. I think the issue arises when you try to generalize on both western mixed Jews and caricatures. I never anticipate it to be able to detect plastic surgery crypto jews, but most normalfags don't even know jews are an ethnicity, so even if we are only able to catch the not so white looking ones, its fine.
I like this idea, seems like it would be pretty funny.
Brody Thomas
Fuck off nigger. Do it yourself and come correct rather than making shit threads trying to con others into doing the work for you.