sherjilozair

What's a better solution than Magento?

Are you using Meteor for this website?

The real question is, where do you get this training data from?

Can Vulcan be used for Deep Learning?

The TF program does learn to sovle FizzBuzz. The above code was used to generate data with which to train the network.

It has been popular belief that although deep learning has been successful in a lot of domains, its biggest shortcoming is high sample complexity, i.e. even for simple problems you need tens of thousands of samples. Thus, deep learning was believed to be unsuitable for one-shot (or low-shot) learning, where the model needs to learn from one or a handful of samples per class.

Brenden Lake et al. showed that a bayesian approach outperforms deep learning based methods in this Nature article: http://web.mit.edu/cocosci/Papers/Science-2015-Lake-1332-8.p...

DeepMind took the challenge, and delivered a deep learning based method, which uses no feature engineering and outperforms human-level performance in low-shot learning, thus providing strong evidence against the widely-held notion that deep learning cannot work with small amounts of data.

Part of the problem in writing a deep learning book, is that very little that warrants being in a book, is actually known. The mainstream deep learning academic community welcomes theoretical work but articles on new techniques which beat SOTA are given much more attention than articles on setting up a theoretical structure of the models. Consequently, that line of research is also under-cited and under-developed. In some ways, 'deep learning' is in a different Kuhnian paradigm altogether. So, people who are used to learning about all the intricacies of classical ML models fail to appreciate deep learning, because the metrics with which you would judge a deep learning model are different from the ones you would use to judge a classical ML model. The same applies for the books in different Kuhnian paradigm.

The FSF does not have a monopoly over the word "free". For a lot of people, "free" means as in "free beer", which this book is, in the online format.

IMO, anything that be done with an HMM can now be done with an RNN. The only advantage that an HMM might have is that training it might be faster using cheaper computational resources. But if you have the $$$ to get yourself a GPU or two, this computational advantage disappears for HMMs.

Might be worth noting that it is very likely that Google, Facebook and Baidu did not even participate in this ImageNet. It is widely believed in the vision community that this dataset is at the end of its run, i.e. any subsequent improvements in performance are not due to scientific breakthroughs but through hyperparameter optimization and ensembling.

Well, what I really wanted to post was his quote on "Fear of AI", not really the whole video, which is https://youtu.be/qP9TOX8T-kI?t=3836

I only get 5. I'm happy. And they are probably people who 'liked' the page much before he announced candidacy.

> Allegedly this practice goes on in America, without the consumer's knowledge.

source?

Why do you need a CUDA account? You don't need one to install CUDA.

This right here made me stop as well. I used to keep a box in plain view. That facilitated the thought of "It's right here, I can smoke any time I want. I'll just not smoke right now, maybe later if I still feel like it, the box is going nowhere."

You can't be further from the truth. A vast magntitude of bugs are due to the system running as if nothing has happened operating on garbage data, and giving you garbage results. If your system is robust to garbage data, say neural network training, good luck finding the bug then.

Sorry, but the problem is how to use a pen to draw machine readable shapes. Machines find reading QR codes and OP's triangles easier than reading letters.

Not necessarily. You can overload @ to do some other kind of object combination, which fits your paradigm better. You're not tied to using it for matrix multiplication.

This is Google we're talking about. I think they will spend time to generalize this situation as much as possible.

This is why it isn't 'by default'.

  def a(b=[]):
      b.append(1)
      print b

Now if you do,

  a()
  a()

you might expect to see

  [1]
  [1]

but you will actually see

  [1]
  [1, 1]

The list is not created at function call time, but at function definition time.

Slightly off-topic, but is anyone aware of an similarly capable library for hand-written text recognition, i.e. ICR?

Why not? DNN-HMM and GMM-HMM won;t have done any better even if trained for 2300 hours.

Is there any such collection of advanced Python patterns, aimed at Python programmers with more than 2-3 years of experience?

Because it's not important. What is more important is that the First Science Sequence was finished. In a detailed report, you may find that the battery didn't last as much as they thought it did (perhaps because of the double landings Philae had to make, etc.), but a public-facing article is not the place for it.

Also, there are probably N things that didn't happen as they thought it did. It is better to mention the K important things (K << N), than the N unimportant things that they got wrong.

If Philae, for example, didn't land on the comet at all, then they would have announced something like this: "Philae took a drive-by of the comet 67 p, and has done in-air experiments that would be valuable for us".

This is a reply to multiple sibling comments. There is actually recent work which shows that deep learning methods can also work WITHOUT any reverse signals: http://s.yosinski.com/dan_cownden_presentation.pdf

None of the parties mentioned actually deny the above equivalence. The reason backprop is a popular idea in deep learning is because people started developing continuous models, where the output (and the error) was a continuous and differentiable function of the input and the weights, which allowed chain rule to be used to compute the gradients, which allowed one to use gradient descent methods. This shift from discrete units to continuous units was termed error backpropogation, and not just chain rule.

If you're looking for programming tutorials on deep learning with some theory on the side, I found these perfect: http://www.deeplearning.net/tutorial/

Also,

3. Being different (hipster) makes the audience remember the video, and hence the product, which is the purpose of the video.

This is exactly how Windows' device drivers work. What you are asking for is the same functionality be extended to userland programs.

The primary problem for this is proof generation. Programming paradigms where proofs are easy to generate would be too limiting. So, although we gain security, our flexibility is tarnished.