Deep learning general

Trying to get this shit installed.
https://github.com/taizan/PaintsChainer

Recently finished working on my bachelor's thesis on extreme learning machines for classification. Pretty much the opposite of deep learning. Only one hidden layer, input weights assigned randomly and never changed, "learning" is just reversing a matrix.
And actually it's not as bad as you could guess, got like 77% on a hard benchmark (forest cover from kaggle). And the "learning" takes like a few second for tens of thousands of samples.

This is my first time posting here so please excuse me if this doesn't belong here, but I'm making two things that remotely involve Machine Learning and AI for a [spoiler] video game [/spoiler] I'm working on.

The first is a navigation volume. I'm using an octree (for now) to partition a space and then activate/deactivate nodes based on if their bounding box is colliding with anything. Agents can then use this structure to dynamically navigate environments in true 3D space. This is useful because I can count of extreme vertical movement because my game involves robots that fly all over the place.

The problem with it is that I'm having trouble optimizing the "dynamic" aspect of it. If anyone has any sort of experience with navigation volumes I'd appreciate some good reading material on it because all I can seem to find on it is the same YouTube and stackexchange threads. I can't even find a decent paper on it. I'm also working on converting it to a k-d tree but that is somewhat easy I jsut haven't ogtten around to it fully.

The second thing is a robust animation system similar to the Motion Matching system in Ubisoft's upcoming game For Honor. If you haven't read up on it and you have a passing interest in how animation works, it's a pretty neat piece of technology. This project is a bit of a stretch goal as I'm just one guy working on all this and it is a pretty sophisticated package but I'm still actively researching it.

>>58703502
nice

>>58703766
It's a clusterfuck between python2 and 3 so you have to cludge shit together just to get it to work.

Working on the kaggle lung cancer competition.

>>58703471
Taking ML course on coursera. Thinking about how to solve a kaggle problem.

>>58704081
What have you done so far? Have you made any submissions?

what is google going to do about malicious programs that can recognize captchas through machine learning? Even a 1 in 3 chance is still very good.

I won't tell because all of you are pajeets.

File: UNADJUSTEDNONRAW_thumb_a2d.jpg (13KB, 224x225px) Image search: [Google]

13KB, 224x225px

>>58704455
33% is good.
sure.

>>58703471
watching https://www.youtube.com/watch?v=vq2nnJ4g6N0

>>58704644
Thanks my dude, will watch this later.

>>58704616
A captcha is pretty worthless if a bot can pass it 33% of the time.

>>58704616
humans probably fail them like 75% of the time. 33% is getting pretty close to being indestinguishable from a dumb human.

i never got the weight/bias part

>>58705339
You never got the whole point on nn in other words.

>>58703471
So i haven't taken a paper on AI yet, but I was doing a bit of looking into it.

Has anyone tried using machine learning techniques to look at house prices? I was basically going to look at different variable factors to try to determine whether a house is over or underpriced relative to other houses based on data about the house.

Anyone tried this?

>>58706066
It's literally the first example of logistic regression you encounter when you start learning ML.

>>58706124
*linear/polynomial regression

Do you need an advanced level of math to get into deep learning?

>>58706338
If you only need it to add some predictive feature to something you are developing, then no.
If you want to do anything significant with it then yes.

>t. ML PhD

What's the best way of doing facial recognition without having to write it yourself?

Basically I have a big amount of videos that I want to tag with character names.

OpenCV to extract faces and then...?

>>58704455
Pay you thousands of dollars for it.

>>58706338
It's better if you want to create your own implementation of a NN.
Else, i don't know, just import this python library and wew lad

>>58703661
No one but stuck up fags care that you brought up video games

Man up a little and just talk about what you like on the anime forum

I really enjoy watching videos about genetic algorithms, neural networks,classifiers,etc and I'd like to start doing some simple shit with TensorFlow and other libraries and I have some questions.
What are some things that I should know before getting myself into deep learning? What are some good resources to learn? I've seen videos from that Siraj guy and learned some basics.
Is there any list of simple projects somewhere that I can use to learn?
Why is Python so much used in deep learning / AI? Isn't it slower than c++ for example?
Do you use GPU to train neural networks?

Please no bully, I just have a bit over 1 year of experience with C/C++.

>>58706338
Not really, just things like discrete, statistics, generic college algebra, trig, calc is all its needed.

Thats usually 1st year college stuff.

Those of you who learned ML outside of a formal setting (i.e. on your own) where did you start?

I've had interest in ML on and off but the subject matter has always appeared daunting.

>>58703471
Im slowly getting motivated enough to start learning it, so far no luck.

>>58708200
Andrew NG course on Coursera is pretty much universally agreed upon to be the first step.
Once you finish that, you won't need to ask here anymore.

>>58703816
>clusterfuck between python2 and 3
this is why i absolutely hate python

>>58704455
>fucking jewgle
>malicious
Pick one.