Deep learning

http://neuralnetworksanddeeplearning.com/chap6.html

by joeyespo • 11 years ago

361 35 11 years ago

billconan• 11 years ago

This is the best neural network tutorial out there. I have been waiting for the missing deep learning chapter for so long, and it finally comes!

my reading for today, thanks for sharing!

jcr• 11 years ago

As a different chapter, this is not exactly a dupe, but it's not the first time links to parts of this book have been posted. Over the last two years, there have been a lot of HN discussions on the various chapters of this book. Here are the ones with comments:

16 days ago - https://news.ycombinator.com/item?id=9863832

8 months ago - https://news.ycombinator.com/item?id=8719371

a year ago - https://news.ycombinator.com/item?id=8258652

a year ago - https://news.ycombinator.com/item?id=8120670

a year ago - https://news.ycombinator.com/item?id=7920183

a year ago - https://news.ycombinator.com/item?id=7588158

two years ago - https://news.ycombinator.com/item?id=6794308

zo1• 11 years ago

How did you do that? Genuinely curious.

film42• 11 years ago

HN search: https://hn.algolia.com/?query=neuralnetworksanddeeplearning....

idunning• 11 years ago

I was really impressed that the author included this caveat:

A word on procedure: In this section, we've smoothly moved from single hidden-layer shallow networks to many-layer convolutional networks. It's all seemed so easy! We make a change and, for the most part, we get an improvement. If you start experimenting, I can guarantee things won't always be so smooth. The reason is that I've presented a cleaned-up narrative, omitting many experiments - including many failed experiments. This cleaned-up narrative will hopefully help you get clear on the basic ideas. But it also runs the risk of conveying an incomplete impression. Getting a good, working network can involve a lot of trial and error, and occasional frustration. In practice, you should expect to engage in quite a bit of experimentation.

There is a lot of "magical thinking" amongst people not actively doing research in the area (and maybe a bit within that community too), and I think it at least partly stems from mainly seeing very successful nets, and never seeing the many failed ideas before those network structures and hyperparameters were hit upon - a sampling bias type thing, where you only read about the things that work.

jacek• 11 years ago

Yes, difficulty of finding right hyperparameters is often overlooked. And it is a very frustrating part of creating a model. And methods like grid search just don't work, because of number of parameters to tune and time to train a network.

DavidSJ2• 11 years ago

Actually, random search works a lot better than grid search for hyperparameter optimization. Usually, only a small number of hyperparameters actually matter, the trick is figuring out which ones. Grid search wastes time on irrelevant dimensions.

That said, any sort of hyperparameter optimization is extremely computationally intensive so random search is far from a panacea.

return0• 11 years ago

It's worth reading Nielsen's essay "Will neural networks and deep learning soon lead to artificial intelligence?" which was added today

http://neuralnetworksanddeeplearning.com/chap6.html#AI

sampo• 11 years ago

And his answer is:

I believe that we are several decades (at least) from using deep learning to develop general AI.

goodness• 11 years ago

I think this is a better summary of his conclusions from that same paragraph:

I conclude that, even rather optimistically, it's going to take many, many deep ideas to build an AI.

The appendix linked there doesn't seem to be ready yet though. In any case, I like how this is phrased. I'd like to see some of the hype around deep learning calm down.

thearn4• 11 years ago

I've sort of run adjacent to the field of machine learning in the last few years, but haven't really dove in to the existing literature. This seems to be a pretty interesting overview.

Out of curiosity, do many implementations of convolutional neural networks take advantage of FFT, DCT, or some other fast orthonormal transform to compute the transition between layers, or are the kernel sizes small enough that there isn't a great advantage to that?

agibsonccc• 11 years ago

Most use an im2col/col2im based impl. I've been experimenting with cuda based fft for ours as well though.

Houshalter4• 11 years ago

Facebook actually does something like that: https://research.facebook.com/blog/879898285375829/fair-open...

They have a patent on it but did open sourced the code. They claim it's up to 24x faster than standard. But that is only true for an extreme use case, its only 2x faster on average.

varelse2• 11 years ago

The bigger the convolution, the faster it gets because a convolution in real space is a multiplication in frequency space.

It breaks even at 5x5 or so and gets dramatically better shortly thereafter. However, most of the convolutional nets in use rely on 3x3 convolutions because I guess reasons:

http://arxiv.org/pdf/1409.1556.pdf (all 3x3)

http://www.cs.unc.edu/~wliu/papers/GoogLeNet.pdf (3x3 and 5x5)

There's probably a new Imagenet winner in this somewhere IMO...

buserror• 11 years ago

Thank you for this, even if it might be a repost, I had missed it, and it's very, very interesting.

dang• 11 years ago

Not a repost; apparently this chapter was just released.

MrBra2• 11 years ago

I just wish the code was in Ruby... but having the author released all this material free I don't feel like actually complaining :) it's more of a subtle hint than anything else.. ;)

MBlume• 11 years ago

Read the material, skip the code, write your own code in Ruby, that's what I've been doing in Clojure.

MrBra2• 11 years ago

That would be a highly effective way of learning but not everybody has the time for it... and also it's not just the ready made code that I envy, but also the added value of someone having already picked the relevant, working libraries, tools and so on..

Deep learning

Related Stories