logoalt Hacker News

joezydecotoday at 2:32 AM5 repliesview on HN

How does one "train" an AI with a flood of random toilet pictures and no corresponding medical data to match it with?


Replies

venturecrueltytoday at 5:37 AM

You pay someone in a developing nation $1.00 per day to look at thousands of photos of shit. Like, how do people think Facebook moderation and semantic labeling happen? Cheap labor in places with no labor laws. It was ever thus.

imglorptoday at 2:40 AM

"potty training". Sorry.

Anyway a chemical or biological sensor in the bowl might be more useful.

Optical could be useful if it's doing spectrographic analysis: the color of poo and urine is sometimes informative.

hackernudestoday at 2:40 AM

They probably do clinical trials (or at least something like that) where they get baseline data from participants through other means.

show 1 reply
captainkrtektoday at 2:46 AM

I think the obvious things are:

- Deviation in consistency/texture/color/etc.

- Obvious signs related to the above (eg: diarrhea, dehydration, blood in stool).

Ultimately though, you can get the same results by just looking down yourself and being curious if things look off...

tldr: this feels like literal internet-of-shit IoT stuff.

g-b-rtoday at 3:42 AM

They probably do match it, with data collected from other sources