I think Machine Learning for Image Manipulation on the Librem Phone would be very
We are going to get crappy specs let’s be honest. According to Antutu, the CPU is 10x
less powerful than even cheaper current phones, I don’t expect the camera to be anywhere
close to the quality of any other recent phone. But we all know this and we are not
ordering the Librem for its spec, but for Freedom.
Given that, I think that ML can really compensate for the so-so camera we are going
to get. I understand it might be seen as “cheating”, but if you are a purist of photography
you should just buy a camera as you’re not going to go far with the librem 5 camera anyway.
Also there is a whole lot of different ML algos out there, not all hallucinate pixels.
For example you could have ML contrast/color/lighting automatic adjustments and all
it does is just automatize what you would do by hand on GIMP. Similarly you could use
ML for HDR, you take pics at different expositions, but when combining them, instead of
hard coding arbitrary parameters, you let the ML part find the best weight for each pic.
It’s pure HDR, no information was “hallucinated”, we just have ML to find the best parameters
It is especially important on phone. On desktop maybe we are okay with hand fine-tuning
dozens of parameters silders on GIMP, but on the phone both because I am on the go and
because of the size of the screen, I want quick easy one-click features.
I might be biased because I am a CS student with specialization in ML, but don’t
think that those things are that complicated. I could completely re-implement myself most
of the so-called “AI” (I hate this term) features of the Pixel / Huawei cameras if there are
Now given the spec of the Librem Phone, we are not going to get this in realtime, but we can
take a picture and then process it in background, no need for dedicated ML chip for that, we can
just leverage the integrated GPU.
Concerning Privacy it very sad what ML is used for. Before it was “Give us your data or terrorists will kill your children”, now it’s “Give us your data or science will not progress and we won’t be able to do ML on it”. It’s absolutely horrible how facial recognition (which should be legally banned imo) is used by FB & others, how
NLP algos run on our private conversation to profile us, etc.
But honestly, like for terrorism, it is yet another excuse for taking more private data. Data can be bought from official hand-annotated dataset, can be anonymized and more importantly we can now use Federated Learning to not only preserve privacy but also distribute the cost of computations.