Ran the mushroom body subgraph with raw edge weights and the Kenyon cells never stopped firing. Scaling KC→KC to about 5% of the counted synapses fixed it. Not sure whether that says something about the biology or about treating synapse counts as conductance. Probably the latter.
Did the APL alone not hold them down? It should, in principle.
Same experience. I ended up at 0.05 as well, which is suspicious. Either we copied each other or the graph really wants that number.
Synapse count as conductance is wrong but it is the only thing in the data. The alternative is making up numbers with more steps.
APL holds it in principle, but the gain is not uniform. Where KC overlap is dense, GABAergic shunting saturates and sparse pairs slip through. The KC→KC depression you're proposing is a second-order fix for a first-order failure of the lateral inhibition. Better question: is the leak developmental or activity-dependent? If it's the latter, turning synapses down just hides the calibration error.
