mochomocha

Netflix. As long as you don't compete directly with the business, the company doesn't care (at least that's how I interpreted my contract when I joined a while ago).

I don't think the parent meant that it's always easy to discern good from evil. The same way I don't think you believe in absolute bothsidesism. But in some extreme cases (such as current invasion of Ukraine or Nazism last century), it would be pretty hard to argue that there's not one better side. Switzerland isn't asked to pick sides in every single conflict. Starting with the clearest cut ones would be moral progress.

At Netflix, we're doing a mix of what Dan calls "CPU Pinning and Isolation" (ie, host-level scheduling controlled by user-space logic) [1] and "Oversubscription at the cluster scheduler level" (through a bunch of custom k8s controllers) to avoid placing unhappy neighbors on the same box in the first place, while oversubscribing the machines based on containers usage patterns.

[1]: https://netflixtechblog.com/predictive-cpu-isolation-of-cont...

YMMV. We've seen M$ worth of cloud savings at Netflix doing pinning right. Knowing that the task scheduler is also heavily forked in Google's kernel, I'm ready to bet they've seen order of magnitude higher savings in their own DCs as well.

In a setup with high-level of containers collocation on large ec2 instances, we've seen the opposite behavior at Netflix: default CFS performing badly. We've AB tested our flavor of custom pinning and measured substantial benefits: https://netflixtechblog.com/predictive-cpu-isolation-of-cont...

PMC data at scale is pretty clear: very often, CFS won't do the right thing and will leave bad HT neighbors on the same core, leading to L1 thrashing, or keep a high-level of imbalance between NUMA sockets leading to degraded LLC hit rate.

I don't think it really matters, the goal post will move to LLC attacks. As long as you have caching involved, I wouldn't bet anything can run "safely".

As mentioned below, this is also how the UCs are doing admissions. You can look at the data for yourself here: https://www.universityofcalifornia.edu/infocenter/admissions... You'll find out that there's a ~5-10% acceptance rate per high school.

Yep. And the data that shows it at play is here: https://www.universityofcalifornia.edu/infocenter/admissions...

My statement was as approximate as the one I was replying to. If the original statement had been "you can use FFT to implement some operations involving matrix multiplication faster" I wouldn't have taken issue with it. But saying "FFT is faster than matrix multiplication" has little meaning.

You might be interested in this line of work: https://eng.uber.com/neural-networks-jpeg/

Training straight from DCT coefficients to avoid spending time learning a similar representation in the bottom layers of the net. I've personally toyed with something similar on GANs to gauge the computational benefits of not doing convolutions in the bottom layers of a net but learning directly in a FFT-like compressed space instead.

Yep, you can successfully do it without convolutions. Here is a pointer if you want to dig deeper: https://eng.uber.com/neural-networks-jpeg/ (there's prior art to that, but this one is well written)

I have trained GANs on raw JPEG coefficients with moderate success as a pet project. Read the raw DCT coefficients without decompressing and train in this space. JPEG-decompress the output of the net to reconstruct an image in the pixel space. There are few papers doing similar things for supervised learning tasks iirc

This is analogous to saying "an umbrella is better than a toaster". They are two different things.

I agree that ML tends to put weaker assumptions on the data than classical statistics and that it's a good thing.

However most ML certainly makes distributional assumptions - they are just weaker. When you're learning a huge deep net with an L2 loss on a regression task, you have a parametric conditional gaussian distribution under the hood. It's not because it's overparametrized that there's no distributional assumption. Vanilla autoencoders are also working under a multivariate gaussian setup as well. Most classifiers are trained under a multinomial distribution assumption etc.

And fat-tailed distributions are definitely a thing. It's just less of a concern for the mainstream CV problems on which people apply DL.

Yes... It's a balance. I don't think tests-only admissions are the panacea either to be clear. The key problem is how to introduce some level of subjectivity in the admission process without creating doors wide open for corruption or cottage industries to "prep your application" with semi-fake accomplishments demonstrating your "soft skills".

As a counter-example, I have gone through a relatable experience though not in India but France (our "preparatory classes" system also has similar high competitiveness/high pressure characteristics, to prepare for a few deciding exams). It only lasts for 2 years, but the stakes are somewhat similar from what I understand from the article.

Having lived in the US for a decade now, one of the biggest culture gap I keep and that probably won't ever go away is how it's somewhat accepted here that "letting rich kids cheat/buy their way into Ivy League is not that bad, and look it pays for a new cafeteria".

In the PGO category, there's also this recent proposal: http://lists.llvm.org/pipermail/llvm-dev/2020-April/140763.h...

At compile-time only, people have also been using trained models to derive cost functions for sequences of instructions (as opposed to analytical models which become very difficult to derive these days given the complexity of modern CPU architectures)

Ha I see - I've been training ML models with millions of data points and millions of features on CPUs for a decade now, I didn't know I needed a GPU - I guess I've been doing it all wrong all this time!

CUDA is nowhere near the de facto in the ML industry. Beyond some specific application domains like CV where DL makes sense, most ML still runs on commodity hardware, and I happen to believe this will be the case for a long time once the Gartner Cycle of Hype settles on the Deep Learning craze and everyone runs out of VC money.

Despite what cloud providers and NVIDIA want you to believe, no you most likely don't need a lot of compute power for your industry problem. Basic knowledge of statistics would help though.

pyarrow

Indeed. Having myself spent some time in the "VST lifestyle business" when I was in grad school (was selling a guitar emulation based on physical modelling synthesis), and now working in ML, I think there's no chance for such an approach to hit "mainstream" anytime soon. Even if you do your inference on CPU, most deep learning libraries are designed for throughput, not latency. In a VST plugin environment, you're also only one of the many components requiring computation, so your computational requirements better be low.

This is really cool! This is kind of a cloud CPU version of Universal Audio Apollo which offloads on local DSP chips.

What's the lowest latency achievable in practice? I'd be surprised if you can run anything under 50ms, so it's probably limited to mixing use-cases, which is already really cool because that tends to be the "plugin-heavy" phase.

How does it work in terms of licensing for the plugins that run remotely? I'm not familiar with common licenses used by plugins.

> I don't know what to do with the anger I'm feeling. Any thoughts on where this energy should go?

I'd personally direct it towards the realization that the US in these difficult times looks more like a plutocracy than a healthy society (I live here). Socialism for the rich, "capitalism" for everyone else. Thinking of creative ways to convince your fellow citizens that this is nowhere near "the best" a society can offer to its constituents would be a good use of extra anger / energy.

"Better late than never" stays true, but when you're fighting against an invisible enemy (because of little testing) that invades exponentially quickly, "late" will unfortunately have a high human lives cost associated to it.

Typical algorithms solving combinatorial problems are usually heavily branching, so not so much vectorizable. For example for a branch-and-cut MIP solver, you could run the LP-relaxed problems on GPU, but most of the logic lives in branching and cutting heuristics that are much harder to parallelize.

Why do you believe this is misinformation? All it's trying to do is mobilize people who understand what an exponential is to act.

I don't think that's what this paper is about though:

"This paper investigates how changing the length of the school year, leaving the basic curriculum unchanged, affects learning and subsequent earnings"

In-person curriculum != online curriculum (social interactions, lab classes, PE...)

Short school year != Interrupted school year

> A child can skip 6 months to a year of school and have almost no negative effects on their life outcomes.

Citation needed.

Very interesting. This independence across blocks can presumably be leveraged at decode time for faster decoding though. Surely there must be decoders out there parallelizing over the blocks on multi-cores arch /GPU?

Do you know how Brunsli & Lepton fare when it comes to parallelizability?

How is that different from insurer incentives to deny coverage or to ask crazy copays because insured people can only switch plans once a year and select from a tiny menu of options that their employers have negotiated on their behalf with an oligopolistic industry? And do you honestly prefer the shady debt collectors industry to the current government money collection mechanisms ?!?