Tensorflow and deep learning, without a PhD, Martin Gorner, Google

[@dbiswas]

This is "THE BEST" workshop on Tensor flow. Thank you so much Martin !!!!!

[@KatySei]

Thanks Martin .One of the best lectures I've ever heard.

[@pymzorr]

Many thanks, following a Deep learning course this semester, and this presentation is worth at least 15 hours of lectures.

[@MartinGorner]

You can now run this yourself with a self-paced code lab: codelabs.developers.google.com/codelabs/cloud-tensorflow-mnist Have fun

[@dbiswas]

You are the "BEST" !!!!!

[@LuisFernandoValenzuela]

Martin Görner thank you very much! 👏

[@bejoscha]

Awesome. Thanks so much. (Both for the video and now for the excellent code lab).

[@STIVESification]

awesome video

[@jogatavid]

That was absolutely great! Thanks Martin!

[@milesdavidsmith]

This dude is giving a great talk.

[@arriva1256]

Really pleasent explanation of the intricacies of deep learning ! thanks so much

[@vishnuviswanath25]

awesome talk!

[@maciejswiechowski6064]

I'm glad I stayed with this video even though I do have a PhD (=> the title).

[@mohamadbakhsh374]

You are the best, thanks

[@rl-rc7kb]

From geometrical pov equation L = W.X + b can be interpreted as rotate and translation. Why we need to redefine addition?

[@MartinGorner]

Yes, it is a kind of "rotation". For one image. We want to rotate 100 images with one formula, but keeping b the same for all. We need to replicate b 100 times to make the matrix dimensions comptible. The little trick called "broadcasting add" is to redefine add so that we can still write the "replicate and add" as just "+".

[@Brad_Jacob]

great video I'm still trying to wrap my head around all of this... Could a PID loop be used in conjunction with the learning rate to reduce the number of oscillations on the gradient descent?

[@othmaneelmeslouhi7702]

hi, it's a very very good and clear presentation. But, i tested the source code, always i found 0 % for accuracy. Why?

[@bensibree-paul7289]

So with enough memory could it learn to see?

[@milesdavidsmith]

Google is always amazing me with their absolute excellence in software engineering.

[@BeyondTheBrink]

A bit confused about the chart at 10:52. The "computed probabilities" should sum up to 1, no? They do sum up to above 2 though on the chart. In the real world the input to the xent would have to be softmaxed to sum up to 1, right? Thx

[@MartinGorner]

Well, softmax "normalises" the vector so the guarantee is that the "norm" of the output vector is 1. Using L1, the sum should be 1. You are right. And that is indeed the norm usually used with softmax. Using L2 is OK too and in this case it is the sum of squares that is 1. That is what I used on that slide.

[@BeyondTheBrink]

Awesome, thx for clarifying

[@siaboonleong]

where do get the source data ?

[@bencdavis]

It uses MNIST which ships with tensorflow. Check the intro tutorial on the tensorflow site.

[@ProjectSuperSport]

link to source code please?

[@maksymonufriienko9788]

github.com/martin-gorner/tensorflow-mnist-tutorial other info you can find at the end of video

[@ProjectSuperSport]

Thanks :)

[@Blomiley]

I never thought I would learn so much from a guy in a t-shirt and cuffed jeans. updating model...

[@MartinGorner]

The next video in the series in online: kzbin.info/www/bejne/rJKvYnxod6mSrrs "Tensorflow, deep learning and modern convolutional neural nets". We build from scratch a neural network that can spot airplanes in aerial imagery and also cover recent (TF 1.4) Tensorflow high-level APIs like tf.layers, tf.estimator and the Dataset API. Developers that already know some basics (relu, softmax, dropout, ...) I recommend you start there to see how a real deep model is built using the latest best practices for convnet design.

[@monicaheddneck8190]

spoiler: 99.3%! :D

[@MartinGorner]

I got to 99.5% with batch normalisation :-)

[@MartinGorner]

Here you go: github.com/martin-gorner/tensorflow-mnist-tutorial/blob/master/mnist_4.2_batchnorm_convolutional.py

[@MartinGorner]

and the video proof: kzbin.info/www/bejne/m4iQY6OViLOEsJY max accuracy 99.56%, average accuracy in the last 3000 iterations: 99.52%

[@milesdavidsmith]

I got 101% accuracy

[@MartinGorner]

Slides for this video: goo.gl/pHeXe7 and there is second part as well now, covering recurrent networks and more: kzbin.info/www/bejne/rKKVn6GAacxphJI (from 1h15'30") Questions welcome, here or on Twitter: @martin_gorner