A new application of machine learning looks both clever and practical, as opposed to the more normal properties of being somewhere between privacy, copyright and danger to life. But until you get too excited, you can’t have it.
The true cost of ML applications varies. Many are free to use, which means they’re jeopardizing someone’s paid income somewhere. Speech recognition puts poor people out of work in call centers. “AI” image generators are robbing artists of their income, and “AI” text generators are threatening writers – at least in those few jobs that survived the web and destroyed print journalism.
Applying ML to image compression and decompression appears to be a relatively safe application. Ever since Michael Barnsley invented fractal image compression in 1987, it feels like adding more intelligence to image compression is a brilliant idea.
The new Attention center model does something different: It uses machine learning to try to identify which parts of an image grab a human’s attention first so it can be selective decompress these regions first.
Load the important bits first
If you’re old enough to remember that GIF images gradually appeared line by line as they were downloaded via a dial-up modem, you’ll immediately grasp the appeal. But now it’s more about mobile and wireless connections, the speed of which fluctuates not only wildly but also unpredictably.
The idea is that a low-resolution version of the entire image appears right at the start, and by the time your visual cortex has decided where to put your pupils, that area of the image is already being sharpened. Then, as your attention wanders through the image, the algorithm has guessed where your eyes will go next and next fills in those parts with more detail. Once those parts are fairly sharp, the rest is filled in, the relatively dull bits last.
If it worked well enough, you probably wouldn’t even realize it happened. The illusion would be that a perfectly sharp version appears right at the start. We recommend playing with this demonstration as long as you have a Chrome-based browser and enable its experimental JPEG XL image renderer: go to
jxl and activate it.
The algorithm is described in a post titled “Open Sourcing the Attention Center Model” on Google’s open source blog… and therein lies the irony, and that’s why conditional mode was used in the previous paragraph. Because this feature uses the new JPEG-XL image format — which Google said it would remove from future versions of Chrome back in October.
It would be unjustified and untenably cynical for us to claim that Google is willing to open source the technology because the format is to be removed from Chrome 110, so we won’t do it. ®
https://www.theregister.com/2022/12/02/ml_attention_center_model_freed/ Google Releases Refined ML Image Compression Model • The Register