Upload MODEL_CARD.md with huggingface_hub
Browse files- MODEL_CARD.md +33 -0
MODEL_CARD.md
ADDED
|
@@ -0,0 +1,33 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
# FLYT3 — trained fly-brain artifacts
|
| 2 |
+
|
| 3 |
+
Tic-tac-toe brains trained inside a simulated *Drosophila melanogaster* CNS
|
| 4 |
+
(MaleCNS v1.0: 166,700 neurons, 25,582,938 fixed synapses, Janelia + Google,
|
| 5 |
+
CC-BY). The wiring is never modified; only a linear readout of descending +
|
| 6 |
+
motor-neuron spike rates (REINFORCE) or KC→MBON synapses (dopamine RPE) was
|
| 7 |
+
trained. Play: the web app in the GitHub repo (seanphan/flyt3).
|
| 8 |
+
|
| 9 |
+
## Artifacts
|
| 10 |
+
|
| 11 |
+
| File | Brain | Algorithm | Games | Skill |
|
| 12 |
+
|---|---|---|---|---|
|
| 13 |
+
| `readout.npz` | Tic-tac-toe (serving) | Symmetric self-play REINFORCE + value baseline + entropy | 1,000,448 | 62% win vs random |
|
| 14 |
+
| `mb_weights.npz` | Tic-tac-toe (mushroom body) | Dopamine-gated KC→MBON depression, no gradients (Handler 2019, Bennett 2021) | 1,000,000 | 74% win vs random |
|
| 15 |
+
| `fire/readout.npz` | Aerial fire (FLAME) | Frozen circuit + 64-unit decoder, 3-class | — | 55.2% (chance 33%) |
|
| 16 |
+
| `fire_satellite/readout.npz` | Etkin satellite tiles | Frozen circuit + standardized 64-unit decoder, 5-class | — | **85.7%** (chance 20%) |
|
| 17 |
+
|
| 18 |
+
## Training data
|
| 19 |
+
- Tic-tac-toe: pure self-play (both flies)
|
| 20 |
+
- Aerial: FLAME drone imagery via alarmod/forest_fire (GPL-3)
|
| 21 |
+
- Satellite: Etkin et al. structure-damage tiles via
|
| 22 |
+
kevincluo/structure_wildfire_damage_classification (apache-2.0)
|
| 23 |
+
|
| 24 |
+
## Honest limits
|
| 25 |
+
Neural dynamics, retinal mapping and dopamine gating are engineering choices,
|
| 26 |
+
not validated biology. The fire decoder reads coarse 99-patch retinal samples;
|
| 27 |
+
a 3° tilt broke the fly_ocr original and our tiles are no more forgiving.
|
| 28 |
+
The tic-tac-toe task is solved — 74% vs random is not superhuman.
|
| 29 |
+
|
| 30 |
+
## Credits
|
| 31 |
+
MaleCNS v1.0: HHMI Janelia + Google (CC-BY) · dopamine rule: Handler 2019,
|
| 32 |
+
Bennett 2021 via adonis-singh/TMNF-C (MIT) · recipe: jerryjliu/fly_ocr (MIT),
|
| 33 |
+
nftechie/doomfly (MIT) · listed in cobanov/awesome-fly (pending PR)
|