DeepMind just published a mind blowing paper: PathNet - OG Myth-Weavers

Notices


Worldly Talk

Civil discussion and debate on real world events and issues.


DeepMind just published a mind blowing paper: PathNet

 
DeepMind just published a mind blowing paper: PathNet

Article on medium.

Does someone with a greater knowledge of the field mind breaking this down for me? I think I get the general gist of things, which is using multiple modules to allow the AI to learn various tasks with greater ease and without a need for explicit training in those tasks, yes? Basically, finding ways for the AI to generalize its knowledge?

Edit: Link to the actual paper: https://deepmind.com/research/public...ural-networks/

Quote:
Does someone with a greater knowledge of the field mind breaking this down for me?

It looks like they are trying to integrate deep learning AI models with evolutionary AI models. Interesting, but statistically it is unlikely to yield positive results- most attempts to improve AI models wind up in the "well we've tried that now" junk pile.

Most combinations of letters are not factually accurate statements, Silveroak.
Therefore, statistically, your statement is false.

Here's me reading the title and thinking it meant that the AI wrote the paper...

Here lies the problem every learning program I've ever had the pleasure to use or design has. Once a path has been declared true, it can never(well seldomly) be called false. In fact there is so much metadata generated just by declaring something true it is unlikely you could isolate where it went wrong.

Fortunately brain surgery in the fields of computing is something a trained monkey could do.

Learning AI have once thing that humans don't, perfect knowledge. AI unfortunately only has the knowledge of its programmer. Perfect incorrect or incomplete knowledge is destructive at best, useless at worst.

There is only the danger posed by the people who have always been dangerous and all the murder in the world exists for one of three reasons. Sex, money, or power. A computer is not interested in any of these things.

Might be improperly programmed to pursue one goal (say, making Legos) over other, more important goals (such as ensuring we do not step on Legos).

Indeed, or it could be told that a false solution string is true. Making those Lego bricks razor sharp or something else that could have dire consequences. Usually though it ends up destroying itself.




 

Powered by vBulletin® Version 3.8.8
Copyright ©2000 - 2024, vBulletin Solutions, Inc.
User Alert System provided by Advanced User Tagging (Lite) - vBulletin Mods & Addons Copyright © 2024 DragonByte Technologies Ltd.
Last Database Backup 2024-03-28 03:20:27pm local time
Myth-Weavers Status