this looks good to go!
6 Matching Annotations
- Last 7 days
-
pl.eoinmurray.info pl.eoinmurray.info
-
- Aug 2026
-
pl.eoinmurray.info pl.eoinmurray.info
-
Summary#
Overall I think this is fine, assuming the conditions mirror those in the smaller, preliminary runs that are in the draft manuscript. The only part I have trouble assessing is the recurrent part (TR-05) - see the comment there.
-
Loop conditionsfrozen PING; trainable PING init; trainable zero init; trainable 0.1 initSeparates built-in dynamics from recurrence learned during task trainingTrainable projections𝑊EI and 𝑊IE only in trainable conditionsThe frozen condition is the mechanistic control
I don't properly understand any of these conditions. But the bottom line is that we want to see if trainable recurrent (and FF) weights allow gradient descent to "find" PING over a wide set of initial conditions that includes conditions that are near to PING in a statistaical sense (e.g. weight magnitudes or initial E/I finring rates)
-
-
pl.eoinmurray.info pl.eoinmurray.info
-
select 0.5 to 25 Hz for later variable-rate PING training.
Looks good - proceed!
-
images from the separate official test partition. Training, validation, and test images therefore did not overlap.
It isn't clear that train vs test is what we need, strictly speaking, because the ability of the network to generalise across class is separate from the question of the information carried in the stimulus in the face of spike/shot noise. I'm not overly worried though, because this is more conservative.
-
with 𝐶𝐸=1 nF, 𝑔L,E=0.05 μS,
presumably these parameters are typical for PING net input layer units
-