KenoFischer

14 ms·

91.
by KenoFischer 6y ago
From an implementation perspective of course, so you can get error propagation on the baseline simulator. For the actual ML part, you'd get Measurements complaining that it doesn't know how to propagate bounds through the surrogat
92.
by KenoFischer 6y ago
See here: https://julialang.org/blog/2019/02/julia-entities/
93.
by KenoFischer 6y ago
We are working on improvements to the baseline simulator also (or rather we basically get them for free because they are part of our core simulation engine), so hopefully that should address some of the first order usability issues in SPIC
94.
by KenoFischer 6y ago
Sure, you can always just run the baseline simulation and compare error rates. You can also do fancier analyses to get some rough idea of robustness and maximum error rates (over your particular parameter domain).
95.
by KenoFischer 6y ago
The 1000x goal here specially refers to algorithmic accelerations. Depending on your baseline simulator, there may also be improvements because Julia's simulation capabilities are extremely fast, but those gains are multiplicative, sin
96.
by KenoFischer 6y ago
Our particular project is about analog simulation, though DARPA is also funding surrogate modeling efforts in the digital domain. On analog in particular, one significant impetus is speeding up mixed signal simulations, since digital simula
97.
by KenoFischer 6y ago
Chris has posted the relevant papers in a sibling comment, but let me try to give an intuitive explanation. Say you have some sort of electronic device like an amplifier. It is made out of many individual transistors and parasitics, but its
98.
by KenoFischer 6y ago
Huh, did not expect our press release to end up on HN, but I'm the PI on this project, so AMA, I guess. Will jump into the questions already asked here and provide some context.
99.
by KenoFischer 6y ago
For practical applications you can do better if you allow some information leakage or multiple rounds of communication. People play various tricks. What kind of leakage is acceptable depends on the application.
100.
by KenoFischer 6y ago
I do wish they would let organizations apply earlier in the year (or even better the year before) so the organizers can know whether or not the organization will be accepted. In fact, many students start talking to us even the previous fall
101.
by KenoFischer 6y ago
Oh, you just don't do correctness proofs over the same type system that is used for dispatch (i.e. "the Julia type system"). We already use an expanded lattice internally for reasoning about Julia code, so there's really
102.
by KenoFischer 6y ago
This package basically exposes information that the standard Julia compiler already knows internally (extracted using some hooks I added last year to make the compiler more extensible) and uses for optimization. As such, it has quite good i
103.
by KenoFischer 6y ago
This work is four years old (with development happening before that), so the Julia GPU capabilities probably weren't good enough at the time. If you wanted to do it today, that'd probably be the way to do it, but would need some b
104.
by KenoFischer 6y ago
It's not entirely embarrassingly parallel because photons from nearby light sources interact, so you need to do a parameter exchange. That said, it is true that it doesn't need the low latency fabric that supercomputers traditiona
105.
by KenoFischer 6y ago
Counting the number of galaxies in these image datasets is actually a highly non-trivial, supercomputer-scale problem. I worked on this a few years back! Paper link for those interested: https://arxiv.org/pdf/1801.10277
106.
by KenoFischer 6y ago
Also see the Julia package that makes it acessible with a high level interface and probably one of the easier ways to play with it: https://github.com/wsmoses/Enzyme.jl .
107.
by KenoFischer 6y ago
That's more on the "nice to have" list (at least from my perspective - others may have different opinions).
108.
by KenoFischer 6y ago
I was talking specifically about the ARM ISA extensions. The accelerators are a different question of course. I understand the folks working on the Linux port are trying to reverse the GPU ISA. Apple could of course also just publish that,
109.
by KenoFischer 6y ago
I'm expecting somebody to add them to LLVM soon. I'd talked to folks at Apple about that some time back, but they weren't able to tell me at the time what their plans were for adding it. That said, I am fully expecting them t
110.
by KenoFischer 6y ago
Agreed on error handling actually. It doesn't quite have the feel that it should in a modern language. I think error handling and lifetimes/mutability are two of the top things we're looking at for a fundamental remodel in 2.
111.
Algorithms for Decision Making(algorithmsbook.com)
1 points by KenoFischer 6y ago|0 comments
112.
by KenoFischer 6y ago
Would be curious what the differences in the Moderna vaccine are (at this level - I know the lipids are different).
113.
by KenoFischer 6y ago
Glad it was useful to you!
114.
by KenoFischer 6y ago
I just gave an example, I didn't mean to say it was the exclusive source of info. Every discussion I've seen of what the layout would actually be was always "roughly 10mm^2, pending finalization of the harness". Also, it
115.
by KenoFischer 6y ago
It was talked about all over the place. E.g. see https://youtu.be/EczW2IWdnOM from June around the one hour mark. As I said, I'm very far from an insider.
116.
by KenoFischer 6y ago
I agree the submission process has been somewhat chaotic, but I don't think it has been exceedingly so. The fact that this was gonna be an alpha test of new technology was clear from the start. Taping out with a modern PDK and open sou
117.
Daydreaming of Cotangents (Automatic Differentiation in Julia) [video](youtube.com)
12 points by KenoFischer 6y ago|0 comments
118.
by KenoFischer 6y ago
The differentiable programming framing is quite useful in the scientific simulation space, because it lets you couple traditional physics simulations to either neural surrogates or some sort of bayesian model. Enough of it is happening in t
119.
by KenoFischer 6y ago
You can use FHE to encrypt functions, but the more standard use is to encrypt inputs and evaluate plain text functions (e.g. in a two part sort of protocol, see e.g. [1]). Regardless, the important difference to obfuscation is that once you
120.
by KenoFischer 6y ago
In FHE, you need some sort of key to decrypt the result. In obfuscation the result comes out clear text. General obfuscation is impossible, but iO gives you basically the best possible result which is "given any method of computing the

More ›