search

CNCLgithub / mot

Model implementation for "Adaptive computation as a new mechanism of human attention"
0 stars 0 forks source link

receptive field test notes #66

Closed eivinasbutkus closed 3 years ago

eivinasbutkus commented 3 years ago

I created a scene where the dot is intersected in the middle by the receptive field because I saw some difficulty in tracking in those instances:

1 1 1

I ran 10000 runs for just 1 time step with and without receptive fields. Found that there is a lower score with no receptive fields than with receptive fields:

julia> mean(scores_rf)
-417.71048461963005

julia> mean(scores_no_rf)
-420.6706421941992

So I'm wondering maybe some pixels get dropped in the receptive fields masks or something like that.

iyildirim commented 3 years ago

Interesting -- so you'd say they're basically the same

From: Eivinas Butkus @.> Sent: Friday, May 28, 2021 7:44 PM To: CNCLgithub/mot @.> Cc: Subscribed @.***> Subject: Re: [CNCLgithub/mot] receptive field test notes (#66)

actually, those differences may be non-significant (p=0.19 from my amateur calculation)..

— You are receiving this because you are subscribed to this thread. Reply to this email directly, view it on GitHubhttps://nam12.safelinks.protection.outlook.com/?url=https%3A%2F%2Fgithub.com%2FCNCLgithub%2Fmot%2Fissues%2F66%23issuecomment-850726320&data=04%7C01%7Cilker.yildirim%40yale.edu%7Cc8714e11098440e6c64708d922327e06%7Cdd8cbebb21394df8b4114e3e87abeb5c%7C0%7C0%7C637578422551349564%7CUnknown%7CTWFpbGZsb3d8eyJWIjoiMC4wLjAwMDAiLCJQIjoiV2luMzIiLCJBTiI6Ik1haWwiLCJXVCI6Mn0%3D%7C1000&sdata=6xJ7DEgvG8bhDvPp7dr2NryEJKkggSb1RMZhdckSxZc%3D&reserved=0, or unsubscribehttps://nam12.safelinks.protection.outlook.com/?url=https%3A%2F%2Fgithub.com%2Fnotifications%2Funsubscribe-auth%2FABCNETNYJCJIQJFSZVFDHELTQATETANCNFSM45XVAXFQ&data=04%7C01%7Cilker.yildirim%40yale.edu%7Cc8714e11098440e6c64708d922327e06%7Cdd8cbebb21394df8b4114e3e87abeb5c%7C0%7C0%7C637578422551349564%7CUnknown%7CTWFpbGZsb3d8eyJWIjoiMC4wLjAwMDAiLCJQIjoiV2luMzIiLCJBTiI6Ik1haWwiLCJXVCI6Mn0%3D%7C1000&sdata=JxydmTCQG65tvuAEnmHBic2wUT8hGG11P%2F8lqnTCUB8%3D&reserved=0.

belledon commented 3 years ago

thanks for posting @eivinasbutkus ,

my main concern isn't that the log score is bad but that there is unwanted entropy like we saw with compose!. look at a sensitivity for trial 10 on fixations (or any other trial) with and without receptive fields. I noticed much more spurious attention on objects that where traversing field boundaries.