>want to experiment with neural nets >download tensorflow

Thank Nvidia for being a closed platform, ran by kikes who prefer everything closed-source.

File: Selection_553.png (400KB, 612x542px) Image search: [Google]

400KB, 612x542px

>>60254558

You can always just use ec2 instances, but there's plenty of other cool shit you can do with just a cpu:

Some ideas, from easiest to hardest:
>linear regression on orchid dataset
>implement a 2 layer percepton in numpy, make it learn xor
>Implement a recurrent neural net and backprop through time from scratch in numpy, train it to sum a sequence or predict the next fibonacci number
>Same thing, but with a LSTM cell
>implement backwards+forwards pass of a convolutional layer, dropout and batch normalization
>Deep Q learning in a simple environment like openai gym's cartpole or swing-up

If you do this you will be more competent than ~95% of the plebs I see applying for machine learning internships

>>60255294
>just use ec2 instances
that costs money

>>60255294
Where would you recommend learning these things online anon?

>>60255314
cs231n.github.io.
Look at the lecture slides, do the programming assignments. If you see something you don't understand, google it. If it's too hard, try Andrew Ng's course first.

>>60255380
Thanks

>>60254558
Doing anything meaningful will be slow unless you have access to a supercomputer.

>>60255850
you don't know what you're talking about. Most nips papers could train on a single 1080 in a few days and practical applications usually only require finetuning a pretrained net on a small dataset.

>>60254558
Iktf. I want to build a new PC but I'm somewhat leaning towards and for GPU since they are supposed to be better with their drivers than nvidia right now. But they miss the cuda framework and barely any nn uses opencl.