search

ogmacorp / AOgmaNeo

Arduino and Desktop compatible fast online/incremental machine learning system.
https://ogma.ai/
Other
53 stars 8 forks source link

along what dimension does feedback concat? #6

Closed baileytom closed 2 weeks ago

baileytom commented 1 month ago

im not great at C++ but I've been reading the whitepaper

i know feedback gets concatenated before the next decoder, but is that along the columns? like there is 2x as many columns as the encoder?

also is there a discord for this?

222464 commented 3 weeks ago

Hi, sorry I only see this now!

The feedback is concatenated to the encoder hidden state such that a column in the decoder projects with the same receptive field onto both the encoder hidden state and feedback. In some sense, you can think of it as concatenating across the column dimension, so there are 2 columns per x,y coordinate now. So yes, there are twice as many columns that are input to the decoder compared to the encoder hidden state. Does this help?

There is no discord for AOgmaNeo yet, although I might make one soon!