Ihmehimmeli spiking neural nets

[Jyrki Alakuijala]

This might be of some interest here as AI is related to compression, and people behind this have worked on Pik and JPEG XL. Ihmehimmeli is a brain-inspired way to do artificial neural nets (recurrent/dynamic/temporal/spiking instead of function/tensor/static).

https://ai.googleblog.com/2019/09/project-ihmehimmeli-temporal-coding-in.html

code at https://github.com/google/ihmehimmeli

[SolidComp]

Can this be fruitfully applied to image compression? What's the connection to JPEG XL, if any?

I wonder which approach would be best for multiple sizes or zooms of the same image. Like on Amazon when you roll over a product image and it zooms.

[Jyrki Alakuijala]

Can this be fruitfully applied to image compression? What's the connection to JPEG XL, if any?

Hopefully one day. The spiking neural nets may allow a 2-3 orders of magnitude energy saving in neural computing (decoding) eventually, so they might be able to provide the needed energy budget reduction to neural compression methods.

There are not many technical connections between Ihmehimmeli and JPEG XL. One connection is that both were optimized using external hyperparameter optimization methods.