Don't we already have facial recognition technology that isn't based on AI? why is throwing AI into the mix suddenly a reasonable product? Liability wavers?

▲

dylan604 5 hours ago | parent | next [-]

I think the facial rec systems you're thinking of will recognize faces, but not ID them. They need you to label a face, and then it recognizes that face with a name from there on. Clearview is different in that you can provide it an unknown face and it returns a name. Whether it's just some ML based AI vs an LLM, it's still under the AI umbrella technically.

▲

lazide 5 hours ago | parent [-]

Uh no? Facial recognition to names has been the bread and butter of facial recognition since the beginning. It’s literally the point.

▲

dylan604 5 hours ago | parent [-]

There are plenty of facial rec systems. Thinking of systems like in iOS Photos, or any of the other similar photo library systems. I think pretty much everyone would be freaked out if they started IDing people in your local libraries.

	▲	anigbrowl 3 hours ago \| parent \| next [-]
		Facebook was doing that 10 years ago
	▲	porridgeraisin 5 hours ago \| parent \| prev \| next [-]
		Note that there is no difference in the model or in the training. The only thing needed to convert ios photos into one that IDs people is access to a database mapping name to image. The IDing part is done after the "AI" part, it's just a dot product.
	▲	joering2 4 hours ago \| parent \| prev \| next [-]
		unsure what you mean by starting IDing? Majority business in US does it already, all banks use facial recognition to know who comes through their door (friend who works in IT at Bank of America told me they implemented it cross all Florida branches sometime in 2009), most large chain gas stations as well, so does car rentals, most hotels, etc. I was recently booted out of Mazda Dealership in Florida because 11 years ago in Georgia I sued Toyota Dealership for a lemon sell, and now they both under same ownership and my name came up on "no business" alert when I entered their offices.
	▲	lazide 5 hours ago \| parent \| prev [-]
		Huh? What relevance does that have with the discussion?

▲

porridgeraisin 5 hours ago | parent | prev [-]

After the literal first one which just measured distance between nose and mouth and stuff like that from the 1960s, everything else has been based on AI.

If my memory serves me, we had a PCA and LDA based one in the 90s and then the 2000s we had a lot of hand-woven adaboosts and (non AI)SIFTs. This is where 3D sensors proved useful, and is the basis for all scifi potrayals of facial recognition(a surface depth map drawn on the face).

In the 2010s, when deep learning became feasible, facial recognition as well as all other AI started using an end to end neural network. This is what is used to this day. It is the first iteration pretty much to work flawlessly regardless of lighting, angle and what not. [1]

Note about the terms AI, ML, Signal processing:

In any given era:

- whatever data-fitting/function approximation method is the latest one is typically called AI.

- the previous generation one is called ML

- the really old now boring ones are called signal processing

Sometimes the calling-it-ML stage is skipped.

[1] All data fitting methods are only as good as the data. Most of these were trained on caucasian people initially so many of them were not as good for other people. These days the ones deployed by Google photos and stuff of course works for other races as well, but many models don't.