laarc.io. Looking for AI work. DMs open. ML discord: discordapp.com/invite/x52Xz3…

Seattle, WA
Joined January 2009
#laarc now has a "suggest title" link on the submission page. Submit a story! laarc.io/ Thanks to the lobste.rs/ crew for suggesting this.
2
0
0
20
I'm interested in alternatives to backprop. Apparently there are several. Specifically, I'd like an alternative that updates only a tiny fraction of the parameters rather than making tiny updates to all parameters. Does anyone know of one? Maybe Kanerva: arxiv.org/abs/1804.01756
2
2
0
32
Shawn Presser retweeted
Since it seems like theres a decent amount of interest the idea is pretty simple. Flow both shapes into spheres, do optimal transport on them, blend the point positions based on the transport map, do convex hull on the sphere, move points back towards the rest positions :0
5
4
0
62
Show this thread
4am thoughts: calm chowder sounds very zen. Also, I fell asleep with Twitch TV on, and woke up to "The Old Timey Computer Show": twitch.tv/oldtimeycomputersh… They're playing old ads for the N64, and it's hilarious.
0
0
0
6
If any of you are in a situation like this, please know it’s not normal. I’ve been a junior dev who asked too many questions (though my coworkers tried to convince me otherwise, and were very gracious). But without getting paid? Only for my first year, and only because I was 17.
Replying to @_adam_barker
Shopify code is so freaking nested and hard to understand, and my boss has asked me to stop asking so many questions and do my own research.. without getting paid. Solving shopify bugs takes everything out of me.
0
0
0
6
This is the cutest Youtube comment I've ever seen. And I'm almost not sorry to annoy you with it. So sweet. The context was Dr. Bashir turning 30: invidious.snopyta.org/watch?v=pTOq2nRy…
0
0
0
8
6am thoughts: I wonder if it's possible to solve a 55^3 Rubik's cube without focusing on the edges last. In other words, the common strategy is to fill each face. But then the edges are scrambled, so you have to do the edges last. Can we mix the steps? invidious.snopyta.org/f9smvQ5fc7Q?t=638
0
0
0
3
Shawn Presser retweeted
Why hasn't evolution ever figured out carbon nanotubes? Materials are right there, CNTs have some unusual and useful properties, & (afaik) they're stable in biology's typical 'warm puddle' conditions.
3
1
0
8
Wow. This is cool as heck. I don't even care about the visual quality -- the fact that someone hooked up a wacom tablet in real-time to CLIP+Siren is no small feat. Impressive accomplishment.
Realtime doodling with CLIP+Siren+SSIM.... (with a soundtrack)
3
6
0
36
Apparently they just use OpenCV for the UI. nitter.cc/htoyryla/status/… Perhaps I underestimated how powerful OpenCV's capabilities are. I noticed that a pytorch3d-esque project used it for realtime plotting too, which was neat; you can have a real GUI open while training a model.
Replying to @westis96
No, just opencv... that's my preferred tool. In fact I always find myself converting back and forth between opencv, PIL and pytorch tensor 😀
1
0
0
6
Randomly came across a master animator on youtube. All their videos are silent, black and white, and hilarious. invidious.snopyta.org/watch?v=q3aT6uEg…
1
0
0
4
I like this style of writing. Maybe follow this person. They seem super enthusiastic, and their points are all true. The TC deepfake definitely wasn't from Corridor, though. (He mentions maybe it was.) Corridor probably wouldn't do this without making a video on their channel.
New Tom Cruise deepfake looks amazing (and, of course scary too) and is doing the rounds on Twitter. Quick technical breakdown/explanation👇
Show this thread
0
0
0
0
For my book on classification I’ll be writing about adversarial examples (& misclassification), and for my course this fall I’ll be teaching a bit on GANs,...and sometimes we make connections when we get confused about stuff 😉. Thanks @theshawwn for keeping “DMs open”.
Replying to @theshawwn
Credit to @drscotthawley for pointing out this line of reasoning, which I had never once made the connection between. arxiv.org/abs/1412.1897 Notice that these look not-dissimilar from a collapsed generator. That's what G is trying to do this whole time!
Show this thread
0
1
0
4
Ooh, this is the first elephant I've seen our BigGAN model generate. Neat. Also, animals are recognizable at a glance more often than not, nowadays.
2
2
0
21
Two months later, here's what BigGAN's cock looks like. It's pretty nice, if I do say so myself.
1
0
0
11
Show this thread
This is a pretty crazy thread. "No gradients required"! They use LSTMs to pass messages back and forth. I don't understand it yet, but it definitely caught my eye. If someone wants to try an ELI5 of this, you'd earn a gold star from me. I'll pin it on my fridge and everything.
Excited to share our preprint 'Meta Learning Backpropagation And Improving It' VS-ML meta learns general purpose learning algos that start to rival backprop in their performance. This is achieved purely by weight sharing LSTMs - no grads required. arxiv.org/abs/2012.14905 👇1/7
Show this thread
0
0
0
17
Shawn Presser retweeted
Excited to share our preprint 'Meta Learning Backpropagation And Improving It' VS-ML meta learns general purpose learning algos that start to rival backprop in their performance. This is achieved purely by weight sharing LSTMs - no grads required. arxiv.org/abs/2012.14905 👇1/7
7
73
7
350
Show this thread
Someone emailed me asking how to reproduce our BigGAN model. I replied with lots of details. Here it is, in case anyone else is trying: gist.github.com/shawwn/5c4a9… tldr: -fix conditional_batch_norm: gamma += 1.0 -use evonorm-s0 -use stop loss: if D_loss < 0.2: D_loss = 0.0
1
2
0
21
BigGAN also points out that you can freeze G with no problems: D remains stable, and its loss continues trending towards zero. That raises a question: what happens if you then resume G? Will G have to work extra hard? Will it give up and generate blobs? It can't give up.
0
0
0
2
Show this thread