What does /sci/ know about machine learning? Pic related

File: IMG_0965.jpg (2MB, 2589x2589px) Image search: [Google]

2MB, 2589x2589px

Clear

>>9111458
It's long and hard....no joke.

>>9111458
Help a brainlet out, what the fuck is your pic?
Just a printscreen from a folder on your phone?

File: IMG_0966.jpg (1MB, 2415x2415px) Image search: [Google]

1MB, 2415x2415px

>>9111499
No these are photos I'm taking in real time and occasionally reposting

File: IMG_0956.jpg (134KB, 995x995px) Image search: [Google]

134KB, 995x995px

>>9111499
The gallery type photos are a screenshot of the iOS photos app

File: IMG_0953.png (674KB, 1242x2208px) Image search: [Google]

674KB, 1242x2208px

>>9111509
This is basically accurate tbqh

File: index6.png (476KB, 735x524px) Image search: [Google]

476KB, 735x524px

Help me /sci/. Been trying to make a neural network that draws bounding boxes around genitalia in hentai. I trained a Faster-RCNN-ResNet neural network on 7,683 annotated images, but the results are shit. There are too many false positives (pic related). Furthermore, when there are correct detections on say a vagina, there will be three or so bounding boxes clustered onto of that vagina when I only need one. Plus, the neural network strangely thinks 2 nipples rubbing against each other are genitalia.

Is it a
1. Not enough data problem?
2. How I annotated the data problem? To keep things simple, there was only one label to cover dicks, vaginas, and anuses. I also grouped genitalia together if they belonged to the same person (e.g. if a woman's vagina and anus were visible together, then they went in the same bounding box) or if they were connected (e.g. vagina and dick in the vagina went in the same bounding box).
3. Something else?

For now, I'm thinking about scaling the project back and focusing on dicks only.

>>9111755
You aren't properly normalizing the inputs

>>9111755
>Been trying to make a neural network that draws bounding boxes around genitalia in hentai
...for what purpose?

>>9111779
Fuck me
>>9111804
Eventually build a neural network that decensors hentai

File: 1441042414084.gif (2MB, 500x281px) Image search: [Google]

2MB, 500x281px

>>9111868
kek, that's brilliant

>>9111868
you're doing god's work, anon

>>9111868
this is exactly what neural networks are meant to do.

>>9111755
>>9111868
>>9111887
>>9111955
>>9111989
t. virgin CS

>>9111996
t. salty nip porn bureaucrat

>>9111868
You realize how many more orders of magnitude difficult that goal is than what you're struggling with now? One thing at a time

>>9111868
Don't listen to him! He plans to censor hentai instead! Why would he be teaching the network to detect genitalia instead of detecting censor bars, if that was his true purpose?

File: 1491920094219.jpg (426KB, 902x1992px) Image search: [Google]

426KB, 902x1992px

>>9112544
you have a keen eye.
my true purpose has been revealed!
You may not help me out today, but I will not rest!
maybe I will rephrase my question and post it in an other thread some other day.
eventually someone will answer my question and then the world will be DOOMED

File: clock ticking.jpg (64KB, 500x300px) Image search: [Google]

64KB, 500x300px

>>9112804
nice pic, I had only the first one

>>9111458
Torture in 2017.

You never finish making a model... You reach a point, explain a phenomenon... Then realize you need to tweak something...
> revaluation of the model
> sensitivity analysis
> run insilico experiments
> Make sense of findings...
> Remind yourself... To never tweak the model... But maybe just one last time, because you think the results might be an artifacts of the model.
> Repeat process
> Realize machine learning isn't a threat to humanity, only if your job can be automated. i. e You don't need to interpret the final product.

>>9111755

Umm, maybe use... This over simplified idea.

>ID person
>ID Skin tone
> ID Shadow
> ID increasing body complexity. Large to small.
> Genitalia complex.

Practice identifying easier body parts, before you reach the meat.