Charlie Loyd @vruba@everything.happens.horse

It’s pretty old-fashioned now. The vanilla DWT, the Tudor-era entropy coder – with only admiration for what it was in 2003, no one would propose a new format like ICER today. But it’s interesting in itself and as a what-if. (What if JPEG 2000 were designed for normal people, basically.)

I was just saying to a friend that JPEG 2000 never succeeded in supplanting JPEG outside a few high-end niches (mapping, archives, medical imaging) due to a scope-crept, complex, IP-unclear, and generally unappealing-to-nonspecialist-developers design, but I think there’s an alternate history not too far away in the multiverse where someone adapted ICER into an OS and browser-supported JPEG replacement in, like, 2005.

https://github.com/TheRealOrange/icer_compression is pretty neat – a mostly complete implementation of ICER, a wavelet-based, error-resilient image compression format designed by NASA 20+ years ago for Spirit and Opportunity.

Just had a startling experience on the internet that left me rattled – and, as I consider it, worried about some of my firmest assumptions. I’ll state it bluntly: a knowledge worker expressed the view that what the world needs is more of is their specific kind of knowledge work.

🧵 1/173

Small elaborations:

A. Tying #1 and #2 together: I think many researchers are building badly shift-variant/asymmetric/anisotropic models without knowing it. If they’re lucky, they make it up by overdoing training.

B. I don’t understood the vision Mamba models well enough to have more than a handful of surface-level takes about them. I hope they overthrow QKV but wouldn’t bet either way.

Overall, I think the subfield would benefit from more geometrical thinking.

@secretasianman Their podcasters are terrible.

Based on informal tinkering and reading. Happy to discuss, within reason.

1. QKV attention transformers work but are gross – O(n²) and equivariance problems – and will not last.

2. Most – not all! – learned convolutions are wasted and should be replaced by fixed bases or frames, or at least grouped convolutions.

3. Diffusion and flow training approaches are immature but way more elegant than x→y training.

5. MMA regularization is good.

One of my favorite things on the Internet is when you watch a “Columbo” clip and there’s a comment where someone describes the concept of Columbo at a second-grade level and is massively upvoted for it. This sounds wry and ironical but I do honestly enjoy it.

I’m tinkering with something where shearlets are very clearly The Right Tool For The Job. There is one pytorch library implementing them in what looks like a good way. Its last commit was 3 years ago and it’s not on PyPI; turns out you have to compile it locally against the CUDA SDK. Using it would instantly make my project unportable. I’m not complaining about the author of the library – if anything, I’m grateful to them. But I’m certainly complaining about the situation.

Waking up early, going on a walk, making breakfast, and then spending two hours meditating on how much I wish someone would write some good, differentiable, richly featured wavelet libraries in python.

@kissane It was @grantimatter who introduced me to it. (I haven’t been keeping up lately but I’m looking forward to catching up eventually.)

Importing libraries: 0.1 seconds
Data loading/decoding: 0.25 seconds
Array manipulation: 0.03 seconds
Optical flow algorithm that does not work: 475 seconds
Writing data: 0.35 seconds

someone who is good at optimization please help me

Like if you see something real and find yourself wanting to think or say “Haha, just like in the movies/books/longform fiction podcasts”, interrogate that impulse at least medium-hard.

A good behavioral rule of thumb, in my experience, is to avoid reflexively comparing real life to entertainment.