shoyer

12 ms·

61.
by shoyer 7y ago
A very similar exchange with T-A happened back in the comment thread from last year: https://news.ycombinator.com/item?id=17432975 As you and several others pointed out last year, the energy considerations here are good eno
62.
by shoyer 7y ago
JAX (github.com/google/jax), which is being developed by many of the authors of Autograd, is a probably a better comparison. At the cost of requiring you to rewrite control flow in a functional way, it eliminates Python's ove
63.
by shoyer 7y ago
XLA is a compiler for array code. It doesn't come with AD -- you need a wrapper like TF or JAX (or Swift, I guess) for that.
64.
by shoyer 7y ago
No kidding. We need to solve climate change now , not whenever strong AI is ready. If we wait several decades (or centuries...) to take action, the effects will be far more severe or possibly intractable. The climate system has memory and
65.
I Got Tenure, but Science Is Still Broken(medium.com)
2 points by shoyer 8y ago|0 comments
66.
by shoyer 8y ago
Here's a link to the actual article in PNAS: https://www.pnas.org/content/early/2019/02/14/1818555116
67.
by shoyer 8y ago
How would you define “using neural networks to solve a given ODE”? I’ll certainly agree that it doesn’t make sense to use a single neural net like function to model the full solution of an ODE. But the entire power of deep learning is that
68.
by shoyer 8y ago
I think it's premature to write off using neural networks to accelerate solving ODEs based on this study. There are quite a few ways to formulate the problem and there's been some promising work in this area recently, e.g., using
69.
by shoyer 8y ago
It is indeed much more efficient. In general, you can evaluate a scalar function and its gradient with less than twice the effort required to compute the scalar function alone -- regardless of the number of parameters.
70.
by shoyer 8y ago
Google is investing in fusion energy: https://ai.google/stories/applied-science/
71.
by shoyer 8y ago
This article misses the most obvious way to visualize this graph: put it on a map!
72.
by shoyer 8y ago
PyTroll [1] (the parent org of SatPy) is one of my favorite examples of a project enabled by the open source model. There are only a handful of folks in each Scandinavian meteorological office that write software for processing satellite da
73.
by shoyer 8y ago
Strongly agreed. I'd love to see a paper a backing up this claim but I'm pretty sure it's just wrong. Likewise, I would be quite surprised if Tesla is really pushing state of the art for the size of their vision models with w
74.
by shoyer 8y ago
Static typing for catching errors is only a small part of the vision for Swift on TensorFlow. The real advantage of static typing is that it enables the compiler to reason to about your code, e.g., to automatically rewrite it for a hardware
75.
by shoyer 8y ago
+1 for Mercurial support. I suspect there are enough Google and Facebook engineers using Mercurial to justify basic support :).
76.
by shoyer 8y ago
i don’t think it’s so much that people become “hot” as people tend to work on the same thing for a while. So of course your biggest hits will be in close proximity to your next biggest hits — they are variations on the same theme! Hot topic
77.
by shoyer 8y ago
Minor correction: the Boston Dynamics robots don't use machine learning. This is a common misconception, but almost all robots today still use hardwired control laws. We don't really know how to make robots that can teach themselv
78.
by shoyer 8y ago
This paper is misleading in calling their method a "deep neural network". It's only "deep" in the sense that it involves stacking multiple physical layers, not in the machine learning sense. From a neural network po
79.
by shoyer 8y ago
A couple of publicly available alternatives I'm aware of include: - Girder: http://girder.readthedocs.io - Intake: https://github.com/ContinuumIO/intake I haven't used any of these, but I agree th
80.
by shoyer 8y ago
Yes, I loved this book! It’s compact enough that you can really get a flavor of things very quickly.
81.
by shoyer 8y ago
> I am almost sure that both would be faster than Numpy; that would be a good comparison. This seems a little unfair :). I’m pretty sure that if you’re using NumPy linked against MKL, it would be exactly as fast as Neanderthal running on
82.
by shoyer 8y ago
If you look at Google's Faculty Research Award page ( https://ai.google/research/outreach/faculty-research-awards/ ), the awards are described as "unrestricted gifts." I'm not sure how it ge
83.
by shoyer 8y ago
The plan of record is already to build pymc4 on top of TensorFlow: https://medium.com/@pymc_devs/theano-tensorflow-and-the-futu...
84.
by shoyer 8y ago
Less fully-featured, sure, but also a bit more cleanly designed (in my opinion). Are there features from attrs that you would miss with dataclasses?
85.
by shoyer 9y ago
This isn’t a joke.
86.
by shoyer 9y ago
No, plasma physics is one of the few areas of modern physics in which quantum mechanics is not relevant: https://www.reddit.com/r/AskPhysics/comments/3nohxy/plasma_p...
87.
by shoyer 9y ago
Note that SciPy wraps all of these in nice Python wrappers. (That's basically its reason for being.)
88.
by shoyer 9y ago
One option is to train a GAN to do domain adaptation. Google (disclaimer: I work there) recently did that very successfully for training robots to grasp: https://research.googleblog.com/2017/10/closing-simulation-t
89.
by shoyer 9y ago
> Farmer's who don't use their seeds are sued for patent infringement if a neighbors seeds blow onto their land and they find out about it. This is not true. This NPR article, written by the same author as top link, was already
90.
by shoyer 9y ago
> each square foot of your research is ~$420/month for rent I think you missed a decimal point or two here -- $42,000/month for a 100 square foot office can't be right :).

More ›