shoyer

10 ms·

31.
by shoyer 6y ago
Not being a “conscious racist” doesn’t make you any less racist. Racism isn’t about intentions, it’s about consequences. It’s a convenient slight of to pretend otherwise.
32.
by shoyer 6y ago
This is not a great idea, unless you’re sure the specific expertise you are building will get you a job. In tech, we typically hire PhDs and postdocs in at the same level. Sure, you’ll probably advance a little faster with experience from
33.
by shoyer 6y ago
Well, the Camp Fire in 2018 caused $16.5B in damages: https://en.wikipedia.org/wiki/Camp_Fire_(2018) And more than $1.6B has been spent on fire suppression in California this year alone: https://en.wikiped
34.
by shoyer 6y ago
What makes you think that (1) that it’s even possible to create such beings, (2) that these would even be good disease models or (3) a lack of model organisms is the main factor holding back medical breakthroughs? None of these is at all ob
35.
by shoyer 6y ago
This is actively dangerous misinformation. In places where the pandemic is not under control, there have been vast numbers of excess deaths: https://www.nytimes.com/interactive/2020/04/21/world/coron
36.
by shoyer 6y ago
Surprisingly Azure pipelines, which I understand is used underneath GitHub actions, does support report CI results in this nice way on GitHub, e.g., see these results from a recent PR on one of my projects: https://github.com&#x2
37.
by shoyer 6y ago
This gets rid of temporary arrays, but this still isn't optimal if z is large. Memory locality means it's faster to apply a scalar operation like z 2+c in a single pass, rather than in two separate passes. Explicitly unrolling lo
38.
by shoyer 6y ago
Unnecessary temporary arrays is definitely a major source of inefficiency when working with NumPy, but recent versions of NumPy go to heroic lengths (via Python reference counting) to avoid doing so in many cases: https://github.
39.
by shoyer 7y ago
Right, you still need to write derivative rules by hand for the primitive operations of an auto-diff system. Automatic differentiation provides composition, it doesn't solve the root mathematical problem of differentiating operations
40.
by shoyer 7y ago
The Hangouts video meetings you have access to do in public Gmail/Calendar are not the same product as "Hangouts Meet", which is only available with a G-Suite account. Meet is definitely a more polished experience.
41.
by shoyer 7y ago
Google Maps does this (at least on iOS), and it's actually pretty convenient. As soon as I click on the search bar, I can immediately click on my copied text to go there without doing the "tap and hold" dance.
42.
by shoyer 7y ago
For a slightly more balanced perspective, try this paper: https://www.sciencedirect.com/science/article/pii/S016635422... : Highlights from that paper (all of which look quite reasonable to me, as a non-expert
43.
by shoyer 7y ago
You can enforce exact symmetry in neural networks with the right sort of model structure. For permutation invariance in particular, take a look at Deep Sets: https://arxiv.org/abs/1703.06114
44.
by shoyer 7y ago
No, in this paper we enforced the Lagrangian structure is enforced as a hard constraint, via the model architecture. It's not just a loss term. Soft constraints via loss functions can also help, but in my experience they are much less
45.
by shoyer 7y ago
What doesn't make sense? You can't solve fluid dynamics (in the vast majority of cases) with a computer algebra system. So if you want to compute forward sensitivities, automatic differentiation is a very reasonable way to do it.
46.
by shoyer 7y ago
JAX still doesn't have an "official" neural network library yet. Source: I work closely with the JAX team at Google.
47.
by shoyer 7y ago
Colab uses real Jupyter notebooks. You can download and run them in Jupyter just fine, I do it all the time. There are a few proprietary extensions like forms (saved as comments in Python code) or colab only magics, but they’re totally opti
48.
by shoyer 7y ago
It's actually extraordinarily difficult to get a DNN model to accurately predict planetary motion over arbitrary time horizons. RNNs have a tendency to either decay or blow-up. The only way I know to make neural nets work well for this
49.
by shoyer 7y ago
> I think you missed the point. Navier-Stokes does not explain Lift, nor does Bernoulli's law or Netwon's third. Like the friction under an ice skate, it's a phenomenon we take for granted in science and engineering withou
50.
by shoyer 7y ago
I think TPUs are great, but I don't understand what you mean by saying they can support "300 GB backprop" on a single device. A TPU v3 has 16 GB of high-bandwidth memory per TPU core: https://cloud.google.com/
51.
by shoyer 7y ago
Most of those disclaimers were added tonight, after seeing my comment :) https://github.com/firmai/pandapy/commit/692c968771bb19d4f12...
52.
by shoyer 7y ago
It's a lovely idea to build pandas like functionality on top of NumPy's structured dtypes, but these benchmarks comparing PandaPy to Pandas are extremely misleading. The largest input dataset has 1258 rows and 9 columns, so basica
53.
by shoyer 7y ago
Employee non-solicitation is in fact enforceable in California, e.g., see https://hkm.com/los-angeles/non-competes/non-solicitation-em...
54.
by shoyer 7y ago
I think it's rather premature to say that the scientific Python stack is adopting Numba. None of the core projects like SciPy, pandas, and scikit-learn have been willing to swap out Cython for Numba. Cython is still dominant and I don&
55.
by shoyer 7y ago
There is undoubtedly some value in this data, but it is far less than you think. Despite the hype about big data, it’s hard to draw firm conclusions from any uncontrolled observational trial. Consider how often nutrition studies contradict
56.
by shoyer 7y ago
The practical counterpoint to Dex is Jax, by many of the same authors: https://github.com/google/jax I’m optimistic that we’ll eventually be able to figure out how to build practical tools for deep learning (and scient
57.
by shoyer 7y ago
Agreed, this paper did not show any meaningful generalization and is sorely short on baselines. If you wanted to train a neural network model that would generalize, you would need build in physical constraints into the model architecture. F
58.
by shoyer 7y ago
I worked for The Climate Corporation when it was acquired by Monsanto (now Bayer) for $1B. There is certainly plenty of potential for innovative tech in agriculture, but it’s a really tough space. Machine learning and big data are not magic
59.
by shoyer 7y ago
I’d be careful before following the MIT media lab designs too carefully. Their agriculture project has been called “mostly smoke and mirrors”: https://www.businessinsider.com/mit-media-lab-personal-food-...
60.
by shoyer 7y ago
> see e.g. lax, which is include in JAX. I think jleber means stax here, for pre-defined NN layers.

More ›