So, Sup Forums, what have you been training your neural network to do?
So, Sup Forums, what have you been training your neural network to do?
Other urls found in this thread:
Self bump even though you can't self bump.
Here's a helicopter that learned how to fly by itself using a neural net.
youtube.com
Video doesn't load for me but I'll give ya a bump, neutral networks are chill. I'm excited to see deepmind play SCII
Bump. Hey what are fun ways to get into machine learning? Fun projects to do?
>I'm excited to see deepmind play SCII
When does that happen?
I've always been really curious about neural networks. Any good reads or vids out there?
Andrew Ng's coursera videos on youtube are a good starting point but you will need a lot of time to watch all of it.
It's surprisingly easy to make an image classifier with Tensorflow, you can do it in less than 50 lines of code.
I've already done the Tensorflow tutorial where you make the MNIST classifier. I'm looking for something *fun*
>fun
Hmmm. I'll have to get back to you on that.
Go back and review your linear algebra and multi variant calculus.
Using Pre written programs to do deep learning works up to a point.
You probably can't tell me why you'd use cross entropy over the average squared difference or difference in activation function softmax vs logistic regression
meme recognition and memeposting
all the pepes all the tiem
just before they learn how to kill us all
That's pretty interesting, I would have thought it would be easier to use deepmind to beat a top SCII player than a top Go player, but maybe I just don't know enough about really high level SCII play.
Input = quantum physics and general relativity
Output = Theory of everything
Can deep learning do this?
Not him but my guess would be you'd use one or the other depending on the shape of your data, but I can't say what would work better what data.
you probably can't function in social situations that depend on your ability to entertain and amuse your guests
shitpost of the year
predict record attributes very slowly
No, but it could conceivably do
Input = all possible measurements of a physical system over time
Output = system that can simulate reality down to at least as much detail as the measurements fed in
So a ToE of sorts, but it would be an incomprehensible black box, not an equation.
I don't know much about neural networks but I do know that sc2 is one of the more complex and decision/reaction driven games on the planet. With Go you have quite a lot of time to analyze and understand you enemy's strategy, but with sc2 your enemy could be hiding tech or scouted your attack early or be pushing earlier than you expect or even just cheesing you in the first couple minutes.
If deepmind is given unlimited APM then it will undoubtedly be able to annihilate any human opponent, but it needs to build up an understanding of the game's mechanics and strategies and counters first. There's a lot that goes into a single match of startcraft.
Yeah, I guess there are a lot more variables to deal with. But can't they just let it play games against itself billions of times again and wind up with an unstoppable juggernaut?
Yes but the real question is what restrictions do you place on its inputs? You can obviously make it only input commands to what is physically on the screen but do you constrict how many commands it can input in a single minute to that of a human player? Do you let it select units without a mouse? If you do, then it could conceivably win any match with some basic units and perfect microing of those units.
It can be fun to write your own library too, but I'm looking for a cool project idea to work on. I've seen that video of the guy using NEAT to learn to play a level of Mario. That looks cool and maybe I'll try that.
Aren't neural net discussions always undergrad shitposting?
BURNNN
i've begun programming it to go to bed at the right hour and to eat more thoughtfully. also to be more positive.
oh you meant in computing. i don't have one