What you're describing is a simplified version of gradient descent (tweaking the weights) and online learning (working on one sample at a time).
This version will not get you far, you will just train a model that solves the last math problem you gave it and maybe some others, but it will probably forget the first ones.
There are other similar procedures that train better, but they've been tried and are currently worse than classical SGD with large batches
Not sure how you can compare vscode with TRAMP. A lot of dev work nowdays is done in containers where you install specific versions of dev tools, compilers, etc. Vscode is one click from seamlessly working inside such a container with its dev tools. TRAMP doesn't provide anything like that, right?
* A west-aligned country is at war with a terrorist organization which is part of the Russian-Iranian-North Korean axis of evil.
* A war which the terrorist organization started unprovoked
* The ally country conducts the most precise strike against militant combatants in history (also completely legal by my understanding of international rules of war)
* Your suggestion is that they should've confiscated their walkie talkies instead
Isn't this a circular argument? The question was what "useful" things do these crypto solutions/companies provide, and your example is a product which allows regulated entities to invest in crypto.
"Why is X valuable? Because X allows you to invest in... X."
Also, in this specific example (Arc), is the solution even considered DeFi? There's a centralized list of whitelister entities and you can only participate if you're a customer of these entities. So they're like banks.
The solution you suggest is irrelevant to the issue mentioned in the article. Even if you use np.random.RandomState, or any other "explicit RNG state", that state will still be copied in the fork() call.
The post just stresses that one should be careful when using random states and multiprocessing, so you should either reseed after forking or using multiprocess/multithread-aware RNG API.
Why use autoscaling and not just launch the instance directly from lambda? The run time is short so there's no danger of two instances running in parallel
> Convolutional layers are just simplifications which make training easier. They are priors in the sense that we know a fully connected layer in image applications would just devolve into a convolutional layer anyway, so we might as well start with a convolution layer. That "design" is the prior. But it's not mandatory; the network would still function without that "prior".
As far as I know this is incorrect. Can you point to a paper that shows this? If by "easier to train" you mean that the models do not overfit training data, then that's the whole point of using correct priors / hypothesis classes.
I'm not sure what bugs you in this paper, but the point is that they decouple the prior architecture from the training/optimization mechanism, and that seems interesting.
Sorry, I meant the claim about f and g. Assuming you meant that F(I) should be continuous functions, you can construct an h from f,g to be cont on I and J, no? So it's not an axiom. Just making sure I understand correctly...
The US, France, and other European countries I can actually understand, since these countries performed military operations in Iraq and/or Syria, but I'll be very surprised if Israel is used as a recruiting tool. If you have some sources implying this I'm interested in reading them!
A lot of "may" and "might" in your reply, and absolutely no facts. The facts are that on ISIS' agenda, fighting Palestinian groups such as Hamas has a higher priority than fighting Israel [1][2], yet people in this thread still manage to "lightly" blame the Israelis, the US, basically anyone except those responsible, as if these people are incapable of independent thought.
I think the context you've quoted only further proves my point. This sentence means absolutely nothing to me, it's trivial and devoid of any actionable information. But because the interviewee is a "famous data scientist" it is suddenly important info that needs to be shared with the world?
I'm going to sound very aggressive, but I don't understand the point of this book. They say
> The material in this book is too valuable not to share.
and after reading the sample I feel their definition of valuable doesn't align with mine. It's a "handbook" but the chapters are interviews that don't go in depth into anything. Here's a sample "question"
> Compassion is also critical for designing beautiful and intuitive products, by solving the pain of the user. Is that how you chose to work in product, as the embodiment of
data?
Really? This reads like an onion article about data science.
This comment is very confusing. First of all, the linked paper doesn't state what you claim it states. The authors show equivalence between two specific frameworks of neural networks: SVM-NN and Regularized-NN, and not equivalence between SVM and NN. Generally, SVM and NN are equivalent only in the sense that all discriminative models are equivalent. The kernel trick in SVM requires your embedding to have an "easily" calculable inner product. I'm not an expert, but I think this places strong constraints on the embeddings you can use.
Second of all, SVM does not create any feature space (i.e., embeddings). It just finds a good separator with a maximal margin. Deep NNs, on the other hand, do create features in their hidden layers.
Anyway, even ignoring these issues, I'm not sure I understood your main point.
You probably already know this, but setting up a cross-compiler toolchain is a super annoying task, and is very dependent on the distro you're using. I'm sure the author knows about cross-compiling, but he probably preferred clear instructions that work for any RPi user.
There might be even stronger reasons to not use cross-compilers such as weird bugs or compiler version compatibility issues.
Moxie is a great researcher and WhisperSystems seem serious. However, I don't understand why you claim that TextSecure is designed by cryptographers.
From what I've seen, they use something called the "Axolotl Ratchet", developed by Trevor Perrin. A quick search of his name didn't yield any crypto papers / research by him.
Also, you write "and has been studied in detail for years"
There are no links/references to code/protocol reviews in the WhisperSystems website.
Again, I have the utmost respect for their research, it's just that from the side of a non-crypto-versed user/coder, Telegram and TextSecure look the same.
I've read most of LYAH, and I have no problem understanding mathematical concepts, but man, Haskell people truely don't know how to convey this "awesomeness" they keep talking about.
1. Where do you use automatic differentiation? I've done machine learning, signal processing, etc., but never even heard of it until now. Why should I care? Your pitch should include that (especially when the wikipedia article doesn't really provide real use cases).
2. What's special about lenses? I tried reading http://www.haskellforall.com/2013/05/program-imperatively-us... but there's no summary of what this is about, and from the first paragraphs it seems like a Haskell workaround for setters and immutability. Again, I feel like the community is not pitching these things correctly. People like me start reading, don't understand the point of it all, and give up.
And I can go on... What is the target audience of these features (or Haskell itself)? Is it people like me, or is it more hardware validation engineers, automatic proof system developers, database people?
This version will not get you far, you will just train a model that solves the last math problem you gave it and maybe some others, but it will probably forget the first ones.
There are other similar procedures that train better, but they've been tried and are currently worse than classical SGD with large batches