Conversation

Notices

1. Embed this notice
   iced depresso (icedquinn@blob.cat)'s status on Wednesday, 02-Oct-2024 07:42:45 JST iced depresso
   > liquid state machines
   :blobcatgendou: we meet again
   • Embed this notice
     iced depresso (icedquinn@blob.cat)'s status on Wednesday, 02-Oct-2024 07:45:35 JST iced depresso

     in reply to
     > It assumes that information is stored, not in “attractors” as is usually assumed in recurrent neural networks, but in the activity pattern of all the neurons which feed-back in a sufficiently recurrent and inter-connected network.
     
     :ablobcatnodblink: yes
   • Embed this notice
     iced depresso (icedquinn@blob.cat)'s status on Wednesday, 02-Oct-2024 07:47:42 JST iced depresso

     in reply to
     i'm guessing the signal decay from the synapse dying is probably important. numenta doesn't store it (they only store that a column fired this frame) but it seems implied that the signal's slow fading could probably be part of decision making as well.
     
     not sure. probably would have to run experiments to find out.
   • Embed this notice
     iced depresso (icedquinn@blob.cat)'s status on Wednesday, 02-Oct-2024 07:53:36 JST iced depresso

     in reply to
     > resistance to input noise
     that would make sense actually. if you communicate with binary pulses, all of the decisions involved in making the pulse are consumed and distilled to go or no-go.
   • Embed this notice
     iced depresso (icedquinn@blob.cat)'s status on Wednesday, 02-Oct-2024 08:32:06 JST iced depresso

     in reply to
     no this isn't quite right. they don't develop the 'liquid' they just drop signals in it, swirl them around, and try to detect patterns out of the swirls. an interesting if bizzare concept but the cortex soup does get finagled by the hippocampus
     
     although maybe they are on to something in having a second layer read the hyper-connections (the soup) to produce outputs :what_cirno:
   • Embed this notice
     iced depresso (icedquinn@blob.cat)'s status on Wednesday, 02-Oct-2024 09:01:38 JST iced depresso

     in reply to

     • LucifarGundam
     @lucifargundam reading science today before i run out of cofe and brain.
     
     i think i read about liquid state machines ten years ago. looks like they are still ... useless?
     
     idk some results implied people are using it in finance but it doesn't appear to be AGI material
   • Embed this notice
     LucifarGundam (lucifargundam@qoto.org)'s status on Wednesday, 02-Oct-2024 09:01:45 JST LucifarGundam

     in reply to

     @icedquinn
     Opposites attract

   • Embed this notice
     iced depresso (icedquinn@blob.cat)'s status on Wednesday, 02-Oct-2024 09:02:04 JST iced depresso

     in reply to

     • LucifarGundam
     @lucifargundam i'm trying to make it self-directed :neocat_what: but i'm out in the woods here with some of this shit
   • Embed this notice
     LucifarGundam (lucifargundam@qoto.org)'s status on Wednesday, 02-Oct-2024 09:02:06 JST LucifarGundam

     in reply to

     @icedquinn

     Ah yes, training models to project a specific output