Re: Paris : In Rush For Profits, AI Safety Issues Are Ignored

On Thu, 13 Feb 2025 10:08:48 -0500, DFS wrote:

On 2/13/2025 2:10 AM, rbowman wrote:
 

Give AI enhanced facial recognition to the cops -- won't that be fun.
Enter 'Minority Report'.

 
https://www.youtube.com/watch?v=gRynAK_A8ko

That's not an isolated case. We all look the same to AIs. Technically, we
don't look like anything. There is no 'knowledge' there.

There is a phenomenon called transfer learning. Train a NN on millions of
images, say cats and dogs, it gets good at the job. There is a point where
it has developed a process to handle images in a general form. Rather than
the cat/dog classifier you can swap in a cow/horse classifier and it will
work.

That leads to some interesting processes for picking up wake words like
'Hey Alexa'. First you digitize the audio waveform and process it to
produce a spectrogram, which is an image.

https://en.wikipedia.org/wiki/Spectrogram

Now you're in an area where NNs are really good -- image classification. 
For speech there can be additional tweaking like using the mel scale. The
bottom line is the 'intelligence' is a lot of FFTs and tensor math.