A place for majestic STEMLORD peacocking, as well as memes about the realities of working in a lab.
Rules
This is a science community. We use the Dawkins definition of meme.
Dude on the right is correct that perturbed gradient descent with threshold functions and backprop feedback was implemented before most of us were born.
The current boom is an embarrassingly parallel task meeting an architecture designed to run that kind of task.
The current boom is an embarrassingly parallel task meeting an architecture designed to run that kind of task.
Plus organizations outside of the FAANGs having hit critical mass on data that's actually useful for mass comparison multiple correlation analyses, and data as a service platforms making things seem sexier to management in those organizations.
Random but why is "embarrassing" or similar adjectives so often used to describe a parallel program? What's embarrassing about it?
"Embarrassingly parallelizable" is just the term for a process that can be perfectly paralleled.
rather odd choice of adjective though
I think the usage implies it's so easy to parallelize that any competent programmer should be embarrassed if they weren't running it in parallel. Whereas many classes of problems can be extremely complex or impossible to parallelize, and running them sequentially would be perfectly acceptable.
It's commonly used in some corners of computer science
It's in the same spirit as the phrase "an embarrassment of riches". So a bit of an archaic usage.
Man i dont know. I had an introductery lecture into ML and we were told of some kernel stuff, where you look at a space that could be infinite dimensional and that you do some math to project into low dimensional feature space, where your seperation still works because of your kernel function.
That isnt some black box art form, that is clearly black magic.
The reached the right end pretty quickly. One of the reasons I gave up on ML rather fast. Hyperparameter tuning is really, really random.
There is truth in this, but it isn't as true as some people seem to think. it's true that trial and error is a real part of working in ml, but it isn't just luck whether something works or not. We do know why some models work better than others for many tasks, there are some cases in which some mannual hyperparameter tuning is good, there was a lot of progress in the last 50 years, and so on.
Maybe we should try sacrificing a farm animal to ML. If we're getting into the realm of magic, there are established practices going back thousands of years.