Sure, here's a TLDW.
Imagine you're training an AI model. You feed it data and test if it comes up with a good answer. Of course it doesn't do that right away, thats why you.re training it. You have to correct it.
If you correct the model by correcting the errors, you get overcompensation problems. If you correct it on the differences between the errors, you get a much better correction.
The term for that is LOSS. You correct on LOSS in stead of on pure ERROR.
7 Comments
abbadon420@sh.itjust.works · 10 pts · 357d
Took me 17 minutes to get the joke. Good video though.
UnRelatedBurner@sh.itjust.works · 3 pts · 356d
can I get my instant gratification please?
abbadon420@sh.itjust.works · 6 pts · 356d
Sure, here's a TLDW.
Imagine you're training an AI model. You feed it data and test if it comes up with a good answer. Of course it doesn't do that right away, thats why you.re training it. You have to correct it.
If you correct the model by correcting the errors, you get overcompensation problems. If you correct it on the differences between the errors, you get a much better correction.
The term for that is LOSS. You correct on LOSS in stead of on pure ERROR.
UnRelatedBurner@sh.itjust.works · 2 pts · 356d
haha, loss
sniggleboots@europe.pub · 6 pts · 357d
I watched that video mere hours ago!
Septimaeus@infosec.pub · 5 pts · 357d
Gradient descent?It’s loss.AtariDump@lemmy.world · 4 pts · 356d