Digital Drosophila Archive
2.0 KiB
Feedback motif contextualization — read-only review
This is not a new learning stage, intervention or model. Source: six archived calibration odors, one seed 131101, frozen_reference; all cases retained. Late interval 500–1,000 ms. Fractions measure positive arrived weights, not causal importance, g attribution or net excitation after inhibition/APL.
| Target idx / type | Five-node internal positive input | Positive sources | Sources needed for 80% | Base late spikes | omit ALLN | omit ALPN |
|---|---|---|---|---|---|---|
| 1197 / lLN1_bc | 8.96–8.99% | 197–197 | 49–49 | 228–228 | 217–218 | 227–227 |
| 1381 / lLN1_bc | 7.86–7.89% | 214–214 | 51–51 | 227–228 | 227–228 | 227–227 |
| 1511 / lLN1_bc | 7.57–7.60% | 205–205 | 52–52 | 228–228 | 219–221 | 227–228 |
| 82312 / KC | 36.16–36.39% | 8–8 | 4–4 | 60–61 | 60–61 | 0–0 |
| 133666 / VA6_adPN | 10.18–10.22% | 119–119 | 31–31 | 212–214 | 0–0 | 208–210 |
Consequence for the next experiment
The two-cell loop can sustain activity in isolation, but the full-brain watched ALLNs receive most positive input from outside the five-cell motif. Do not target only this pair as the established unique cause. Existing conditional omissions freeze other recorded events: removing ALLN arrivals still leaves PN/other recorded drive. This establishes conditional sufficiency of residual input, NOT that an ALLN knockout in the live brain would fail. Likewise a large input fraction is not a causal fraction; small inputs can matter near threshold. The VA6->KC edge illustrates why fractions alone are not a deletion target ranking. Future separately authorized model tests should first validate operator semantics, then assess distributed local feedback and preservation of stimulus information, before learning. No gain/threshold/eta/mask tuning is licensed by this table. No new simulations, learning, weights or frozen-file changes. No heldout data used. Usable associative learning remains unachieved.