On Thursday, May 27, 2021, at 5:52 AM, John Rose wrote:
> I would consider OCR (Optical Character Recognition) as a compression. The
> characters extracted are lossless and the noise and background is lossily
> compressed out.
Yes but... The compression of text in the Hutter Prize or LTBC is measuring the
intelligence i.e. the pattern finding ability. Recognizing letters on a screen,
then storing them, and compressing them, while dumping the images, is same as
the contest goal....... but your contest idea and the HP/LTCB don't ask vision
to be compressed !, so there's no need to use it in the "test" for an AI. ---
If you do want to use images, you need to compress the images. If no images, no
compress images, and no images either are needed. The only goal is to compress,
and preferably quickly with little memory (HDD, RAM, Cache)
used........................You might say but yes let's use optical
recognition, see if they can *Decompress the same text, if they can't then they
failed image recognition, you can throw away the images and no compress them.
Well, what if you decompress but the compression sucked or is half there, is
the problem text compression or image recognition and by how much error? You
can't test the image recognition error size because you asked us not to
compress/test the AI's ability to recognize patterns in images. If you fail to
recognize a visual letter, it would affect the next nearby predictions of the
next *Text, so you'll get a large file (since you still know the next letter,
simply are trying to use the AI so not to store it in full), or big error if
use Perplexity evaluation, and we can't tell if this is because the prediction
for the text context was poor or because some the letters were not received
well and had half or no context and was simply guessing at best or storing in
full the next letter 0 compression gain. In short, vision recognition would
affect the text prediction score, but you would need to include vision
recognition error, because you wouldn't be able to know if your poor text
predictions were based on texts AI or vision AI abilities.
------------------------------------------
Artificial General Intelligence List: AGI
Permalink:
https://agi.topicbox.com/groups/agi/T95f11a183fb9b6e1-Mbadc2bd16e247bb2ff5db921
Delivery options: https://agi.topicbox.com/groups/agi/subscription
