hey /g/ can we talk machine """"""learning""""""?

File: 1476123204898.png (212KB, 475x450px) Image search: [Google]

212KB, 475x450px

>>58834729
what's the input?

I kinda want to fuck around with image recognition, but all I did so far is building autistic websites for weebs. What would be a good start? Do I just download a working thing or is the standard way to write your own?

Goal would be archived if the program can determine the possible expressions of animu.

Pic related should be tagged happy, smug and maybe eating

>>58834884
good luck you NEET

File: [Ohys-Raws] Girlish Number - 09 (TBS 1280x720 x264 AAC).mp4_snapshot_16.58_[2016.12.02_18.41.14].jpg (141KB, 1023x714px) Image search: [Google]

141KB, 1023x714px

>>58834904
joke's on you, I manage a sprint team of code monkeys and my superior is drowing in shit from the business, so since I'm paid to do some simple shit here and there I have enough time to code from work.
My Environment runs at home, but I also code here if I want to or if nano isn't enough.

At 2k€ after taxes it's fine for being employed here only 8 months.

>>58834884
i have multiple inputs at the moment
speed, direction, position and then a matrix of numbers that describe the surrounding environment
but i'd like to make him be able to adapt new inputs and outputs, maybe another time though

for image processing i'd say convolution neural net with a supervised learning approach
or you could do a neuro-fuzzy (or another kind of adaptive fuzzy approach) for the expression determination

>>58835278
not keen with the lingo but I'd assume i'd give the program a number of samples and then have it guess in the first phase, where i'll provide inputs to the accuracy to make it as "me" as possible.

So for the first stage it could just give me everything it finds with a percentage or something. Since I've only read about the principles so far I'll only start 2018 the earliest on that though.

>>58835485
alright
yeah the idea is to give him a set of training data and answers to them, in this case images of anime facial expressions, and then he will iterate through the list of images (with a large variety of expressions to avoid overfitting to a specific kind) and change the weights for each node depending on the accuracy of his prediction
with enough training the net should be able to differentiate between expressions
i think it would be quite effective if the neural net didnt make the decision, but rather returned fuzzified data of his prediction,, like, im 60% sure she's happy, 15% sure she's eating, and 25% sure she's smug, and then use a fuzzy controller to defuzzify and get the answer
im not that experienced with deep learning nets though, i could be wrong with the whole training stuff regarding convolutional neural nets when it comes to the bias of each neuron, but i'd assume those are still decided by trial and error (i.e. guess, get answer, update weights/biases, guess again on another image)

>>58835642
yeah that's also how I understood the principle when I learned about it for my boss who wanted to sell the knowledge but not learn it for himself.

I basically watched the Videos about LearnFun and PlayFun on Super Mario by some dude on Youtube and figured this is how neural networks should be done.
Computers can't "intelligence" for you, it's all comparing quantified data with some "status".

So I'd definitely go the fuzzy route and give my input to adjust the weightings.

File: output.webm (2MB, 1500x300px) Image search: [Google]

2MB, 1500x300px

i'm working on conv autoencoder, using Tensorflow

>>58834729
working on face recognition, using a network model based on the Facenet paper, paired with some SVM classification

currently aligning the youtube faces database, few 100K images


>>58834884
if you can gather enough labeled data, its shit
easy

but you will need alot of data

>>58834729
>4D brain
>Thinking in 4D

How does that even work? That would imply past, present and future are irrelevant. To you they seem like a stream of events flowing in one direction. You can observe all the events from start to end and intervene at any point in space/time.

What the fuck?

>>58835746
>>58834884
that being said, there are networks and models out there for facial expression detection, but trained with humans obvs, could be trained with animu tho

>>58835702
nice, how are you enjoying tensorflow?

>>58835746
>>58835773
rad, how does it differ in regards to efficiency and accuracy compared to for example using haar-cascades?

>>58835769
don't worry, it's normal your tiny brain cant understand this concept.

>>58835818
I failed maths in high school. Sucks to be a brainlet.

>>58835788
didnt really look into facial expression stuff,
but its probably faster as it just needs a network forward to generate a classification

face recognition, for example openface needs ~58.9 ms on CPU and ~13.72 ms on GPU for a single pass with torch

accuray depends on your training data

File: [Ohys-Raws] Gabriel Dropout - 01 (AT-X 1280x720 x264 AAC).mp4_snapshot_12.27_[2017.01.09_19.30.33].jpg (102KB, 712x719px) Image search: [Google]

102KB, 712x719px

>>58835746
>>58835773
http://animu.date
don't underestimate the meme stash. After 1k guided training iterations I would unleash the thing on /a/ to fetch stuff for users to rate the accuracy, so training data should be no problem at all.

We here at animu.date SAP (shitposting assistance programs) are a megalomaniac bunch of me and Anon which will make Shitposting great again.

>>58835894
>Youkoso

>>58835769
>That would imply past, present and future are irrelevant.
> You can observe all the events from start to end

he thinks with our time as being singular and perceives events unchanging
hence the start and end of an event is irrelevant and contradictory because the entire event is already existent as a whole, observable from every given direction from his 4D perspective

>>58835916
What about higher dimensions? M-Theory goes all the way to 11 dimensions. How do I even perceive that?

>>58835859
not bad, the issue i encountered with haar-cascades and template comparison for face/iris detection was the amount of data that needed to be stored
i'd assume its less with a neural net

>>58835951
don't know, im a brainlet past 4 dimensions

Still scraping and images of boorus to train my cute girl neural net that I will use to sort them into folders.

>>58835702
Captcha guessing?

>>58835788
>nice, how are you enjoying tensorflow?
it's nice, bit pain but it still better than using caffe

>>58836024
try to understand the image and reproduce it

>>58834884
Smug? More like content.

>>58835894
>http://animu.date
Consider this :
>install browser plugin
>plugin reads 4chan post you're about to send
>apply sentiment analysis on said post
>recommends animu girl reaction pic based on result

pls r8 my idea

>>58836040
>>58836027
alright, ill probably try it out next time, caffe was shit
would you be able to produce an image without perceiving it?

File: increasingly nervous anime girl on the phone.jpg (169KB, 785x558px) Image search: [Google]

169KB, 785x558px

bump
anyone still here?

>>58837460
Where do I start with machine learning? I know programming but no idea on machine learning or how it works

ML is a fad.

>>58837983
start simple
machine learning and ai is honestly a meme, but an neat one to keep yourself occupied with
do something that you could be able to do with a series of if-statements
like make a dot follow your cursor or a dot that avoids your cursor