Hacker News Reader

Comment by TeMPOraL | Hacker News Reader

Damn, there goes my idea. The other day I wondered how far you could go by embedding classical and neural filters in plastic sheets, to make a "magic lens" that does some img2img transformation in real time. E.g. pixelizing should be relatively simple, but I wonder if you could go as far as transforming a small image classifier (say, "cat" vs. "hot dog") into a physical object, to create a passive "lens" that turns incoming light into words ("cat", "hot dog", "N/A")? Feels like it should be doable.

Anyway, thanks for the link; I'll study the paper and look up at the references.

View on HN

> pixelizing should be relatively simple

https://www.yankodesign.com/2024/07/04/this-crystal-fragment...:

“The Pixel Mirror measures 16mm x 16mm x 10mm, which means it’s small enough to be worn as a pendant in a necklace. Monoli’s series of wearable and handheld prisms are all handmade, and because of the nature of polishing natural stones, they are not perfect square “pixels”. They are handmade to suit the condition of the available stone.”

Discussed on HN in

https://news.ycombinator.com/item?id=40895907 and https://news.ycombinator.com/item?id=40890633

I've had the same idea for a while, but the classifier I had in mind was for identifying wood ("yep, it's wood"). I wonder if this kind of tech could be used for low-light enhancement (night vision). This would require "injecting" light into the "lens" I guess.

MNIST classifier using a DNN: https://www.nature.com/articles/s44172-024-00211-6

Some other interesting things in the "capture more with clever hardware" bucket:

* https://en.wikipedia.org/wiki/Light_field_camera

* https://en.wikipedia.org/wiki/Angle-sensitive_pixel

Also, something I dimly remembered about using hard-to-forge glass spheres as physical security tokens, taking advantage of their inner structure:

* https://news.mit.edu/2002/gershenfeld-0925

> Also, something I dimly remembered about using hard-to-forge glass spheres as physical security tokens, taking advantage of their inner structure

Cool! This made me remember of another, simpler technique I saw somewhere: take a bunch of different colors of sprinkles[0], mix them and sprinkle on something sticky. You now have a 2D barcode that's guaranteed (statistically) to be unique. It's also reasonably easy for a computer to scan quickly, with low error rates. Perfect for tracking individual items produced in mass quantity - in fact, I first heard about this in context of tagging individual pills in pharmaceutical products, for purposes of quality control and audit trail. The principle generalizes to confetti, colored beads, etc. of all sizes, but of course using bakery supplies gives you a food-grade stamp that you can also eat :).

I think we pay too little attention to the analog world these days; plenty of fascinating inventions can be made by applying computer technology to shift complexity from code into structure of physical objects.

[0] - https://en.wikipedia.org/wiki/Sprinkles

> but of course using bakery supplies gives you a food-grade stamp that you can also eat :).

This requires that the "something sticky" also be edible. Hmm... choices...

"For verification, we'll need something you know, and something you own. Simply come over here to the kiosk with the keypad and the scanner to... Wait... what did you do with your Security Cupcake?"

I suppose if your lens is actually a vertical stack of lenses, each with hundreds of inputs and outputs, then why wouldn't this work? Although I cannot fathom fabricating it. Maybe start by finding the absolute simplest/smallest image classifier you can

EDIT: Of course people are already doing this. D2NN seems to be the keyword, Diffractive Deep Neural Networks

Yes, by "lens" I meant a possibly large stack of optical elements, each designed to perform a single computational step. Kind of like layers of an NN, but etched on sheets of plastic.

Thanks for finding the proper term for this!

Hardware neural nets. Super cool idea.

Reply on news.ycombinator.com