Audit frozen LIF semantics and real recurrent motifs without learning changes
This commit is contained in:
Digital Drosophila Archive
@@ -0,0 +1,71 @@
|
||||
# Phase 4AC：反馈环路足以持续放电，尚未得到学习修复
|
||||
|
||||
状态：`COMPLETE-SEMANTICS-AUDIT / LEARNING-NOT-ACHIEVED`。
|
||||
|
||||
## 这次真正增加的证据
|
||||
|
||||
从已有异常传播路径选取真实 MaleCNS 神经元和 13 条真实连接。
|
||||
只施加一次初始脉冲，随后无外部输入，观察 1 s。所有真实回路测试继续使用冻结动力学。
|
||||
|
||||
| 小回路 | 总放电数 | 500–1,000 ms 放电数 |
|
||||
|---|---:|---:|
|
||||
| 两枚 ALLN 互相连接 | 473 | 238 |
|
||||
| 三枚 ALLN | 1,064 | 535 |
|
||||
| 三枚 ALLN + VA6 PN + KC | 1,483 | 747 |
|
||||
| 只保留四条真实前向连接的对照 | 3 | 0 |
|
||||
| 五神经元回路，无初始输入 | 0 | 0 |
|
||||
|
||||
最小互相连接的两枚神经元是：
|
||||
|
||||
- neuron_idx 1381 / body_id 11478 → neuron_idx 1511 / body_id 11620：235 contacts。
|
||||
- neuron_idx 1511 / body_id 11620 → neuron_idx 1381 / body_id 11478：283 contacts。
|
||||
- 两者 metadata 均为 ALLN / lLN1_bc，递质 annotation 和 ground truth 均为 acetylcholine。未改变符号。
|
||||
|
||||
它们在冻结 LIF 下，从一次输入即可相互维持放电到 1 s 观察期末。
|
||||
五神经元回路的 VA6 PN（133666 / 527577）有 297 次放电，KC（82312 / 109321）有 80 次放电。
|
||||
这把“持续反馈驱动 KC”缩小到可逐步检查的真实局部回路。
|
||||
|
||||
**限定结论：**这是隔离回路中的充分性证据，不是全脑必要性证据；外部抑制、分级 APL 及其他回路在这里不参与。
|
||||
前向对照移除了反馈连接，只是诊断 harness，不是提交了一个删边新模型。1 s 持续不等于证明无限持续。
|
||||
|
||||
## 原作者语义差异，不等于已找到修复
|
||||
|
||||
[Shiu 作者固定版本代码](https://github.com/philshiu/Drosophila_brain_model/blob/91bdd1e7dcf193f3e7ca5a8933497fcef63b7960/model.py)
|
||||
在 spike 后清零 g，v/g 的不应期更新受限，使用 linear 积分和严格 > 阈值；刺激细胞的 Poisson 输入及不应期也与本项目不同。
|
||||
本项目的冻结版本明确记录了保留 g、不应期继续衰减/接收输入、Euler 积分等适配选择。
|
||||
因此不能把它说成 CUDA 引入的新错误，也不能把“g 清零”称为严格复现原作者模型。
|
||||
|
||||
本次单细胞 component tests 有意保留冻结的其他数学语义，仅分别比较 g reset 和 refractory gate。
|
||||
真实 13 条连接强度加一个单突触 fixture 下，冻结策略、reset-only、reset+gate 的单次输入 spike 数相同（0 或 1）。
|
||||
单独冻结不应期 g、但不清零 g，反而在 283-contact fixture 中产生 2 次 spike，而其他三种策略为 1 次。
|
||||
人工持续输入停止后，冻结策略有 4 次尾随 spike，reset-only 有 1 次、reset+gate 有 0 次，gate-only 则有 14 次。
|
||||
|
||||
这些结果表明组件选择确实会改变响应，但不支持“随便加一个不应期限制就能解决”。
|
||||
两个强连接各自仍能从一次输入触发一次响应，所以也不能仅凭单细胞 reset 测试宣称真实反馈环已被消除。
|
||||
所有替代组件仅运行于人工单细胞 fixture，没有部署到 MaleCNS 全脑或训练实验。
|
||||
|
||||
[Brian2 2.5.1 固定实现](https://github.com/brian-team/brian2/blob/577bb35ec677947d6f3d4af367a8a4b3b88815ee/brian2/groups/neurongroup.py)
|
||||
在默认非 legacy 设置下，不应期是 22 steps；第 23 step 才积分某些新到达输入属于调度顺序，不是不应期延长。
|
||||
原作者模型、license、环境文件、固定 commit、SHA-256 与源码位置均已记录；尚未执行独立 Brian2 oracle。
|
||||
|
||||
## 验证
|
||||
|
||||
- 7 项单元测试通过；17 组单细胞 fixture × 4 策略均有限且重复一致。
|
||||
- 五个真实小回路的全部事件、计数、每步 V/g/ref、最终状态和环形队列 CPU/GPU 精确一致。
|
||||
- 五组 deterministic-v2 全脑参考 trial 的事件、计数、最终 V/g 精确一致。
|
||||
- 逐条核对真实 edge_id、源 CSR 行、目标和原始 count；准备输入另有 SHA-256。
|
||||
- 完整性终检覆盖原快照 230,239 项，以及 Phase 4AA / 4AB 的模型与实验 manifest。
|
||||
- GPU 诊断含构建/观察/回归共 95.398 s；本进程 peak allocated VRAM 285,994,496 bytes。不是性能 benchmark，不含 ComfyUI 驻留显存。
|
||||
- 保留旧加载器只读 NumPy→Torch 警告；未修改加载器，完整性和回归另行验证。
|
||||
|
||||
## 学习目标与下一道关口
|
||||
|
||||
当前没有新的 learned model，没有训练，也没有重新宣称学习 PASS。
|
||||
可用学习仍须在独立评价种子上证明 A 相对中性 B 的特异性、MBON/downstream 功能变化及完整对照；权重变化、稀疏化或静默都不能替代。
|
||||
|
||||
下一步应先在明确授权的新模型分支中验证作者算子与有生理依据的局部反馈机制，并用独立 oracle 区分每一项语义。
|
||||
不能将减少放电自动选为正确模型；还必须保留输入区别、可传播性和后续关联记忆能力。
|
||||
不改 raw、CSR、原始突触数、已定递质符号、冻结版本、APL calibration、eligibility-v2 或 plastic mask。
|
||||
不进入游戏、decoder、RL 或视觉路线。
|
||||
|
||||
Git 只同步源代码、小报告与校验清单。完整数据仍在服务器，不能声称已完成异机大数据备份。
|
||||
@@ -0,0 +1,83 @@
|
||||
# Phase 4AC — LIF semantics and unchanged-dynamics motif audit
|
||||
|
||||
Status: COMPLETE-SEMANTICS-AUDIT / LEARNING-NOT-ACHIEVED.
|
||||
No alternative dynamics deployed to MaleCNS; no conditioning, learned candidate or new formal runtime.
|
||||
|
||||
## Author model versus frozen adapted model
|
||||
|
||||
Pinned author model resets g on a spike and declares v/g `(unless refractory)`, uses linear integration, strict > threshold and voltage Poisson input with zero refractory on stimulated cells.
|
||||
The frozen adapted model retains g, decays/accepts events during refractory, uses explicit Euler and >= threshold, and forced sensory spikes respect refractory.
|
||||
These are documented model choices, not a newly established GPU regression. g is an additive voltage-valued drive, not a measured conductance.
|
||||
W_syn=0.275 is transferred from a free parameter calibrated in the original FlyWire feeding model, not a measured universal MaleCNS synaptic efficacy.
|
||||
|
||||
## Synthetic single-cell component tests
|
||||
|
||||
All impulse amplitudes derive from the complete unique contact counts of the predeclared 13 real motif edges plus a one-contact fixture; other boundary/sustained cases are explicitly synthetic unit tests.
|
||||
These ablations retain frozen Euler arithmetic/scheduling and are NOT complete Brian2 replication. No successful amplitude or policy was selected after seeing results.
|
||||
|
||||
| Single-impulse fixture | Frozen | Reset g only | Refractory gate only | Reset + gate |
|
||||
|---|---:|---:|---:|---:|
|
||||
| one_impulse_contacts_1 | 0 | 0 | 0 | 0 |
|
||||
| one_impulse_contacts_104 | 0 | 0 | 0 | 0 |
|
||||
| one_impulse_contacts_172 | 1 | 1 | 1 | 1 |
|
||||
| one_impulse_contacts_210 | 1 | 1 | 1 | 1 |
|
||||
| one_impulse_contacts_226 | 1 | 1 | 1 | 1 |
|
||||
| one_impulse_contacts_235 | 1 | 1 | 1 | 1 |
|
||||
| one_impulse_contacts_259 | 1 | 1 | 1 | 1 |
|
||||
| one_impulse_contacts_283 | 1 | 1 | 2 | 1 |
|
||||
| one_impulse_contacts_43 | 0 | 0 | 0 | 0 |
|
||||
| one_impulse_contacts_57 | 0 | 0 | 0 | 0 |
|
||||
| one_impulse_contacts_64 | 0 | 0 | 0 | 0 |
|
||||
| one_impulse_contacts_70 | 0 | 0 | 0 | 0 |
|
||||
| one_impulse_contacts_76 | 0 | 0 | 0 | 0 |
|
||||
| one_impulse_contacts_95 | 0 | 0 | 0 | 0 |
|
||||
|
||||
Reset-only and combined policies can produce at most one spike from a single isolated input by construction. This does not establish biological correctness or whole-brain learning improvement.
|
||||
|
||||
## Real anatomical motifs — frozen mathematics ONLY
|
||||
|
||||
One forced seed spike at step 0 except the no-input control; no subsequent input; observed for 1 s. Late window=500–1,000 ms.
|
||||
The recurrent pair/trio/five are induced subgraphs. The feedforward control selects four existing edges and removes feedback only within this isolated diagnostic harness.
|
||||
External inhibition, graded APL and other whole-brain inputs are absent. A sustained isolated motif is evidence of sufficiency in this harness, NOT necessity in the complete brain.
|
||||
|
||||
| Motif | neuron_idx | Total spikes | Late spikes | Last ms |
|
||||
|---|---:|---:|---:|---:|
|
||||
| pair | 1381 | 237 | 119 | 999.3 |
|
||||
| pair | 1511 | 236 | 119 | 997.5 |
|
||||
| trio | 1197 | 355 | 179 | 999.3 |
|
||||
| trio | 1381 | 355 | 178 | 997.8 |
|
||||
| trio | 1511 | 354 | 178 | 997.4 |
|
||||
| five | 1197 | 370 | 185 | 998.8 |
|
||||
| five | 1381 | 369 | 185 | 997.4 |
|
||||
| five | 1511 | 367 | 186 | 999.5 |
|
||||
| five | 82312 | 80 | 41 | 990.3 |
|
||||
| five | 133666 | 297 | 150 | 999.5 |
|
||||
| feedforward | 1197 | 1 | 0 | 10.0 |
|
||||
| feedforward | 1381 | 1 | 0 | 0.0 |
|
||||
| feedforward | 1511 | 1 | 0 | 4.9 |
|
||||
| feedforward | 82312 | 0 | 0 | none |
|
||||
| feedforward | 133666 | 0 | 0 | none |
|
||||
| no_input | 1197 | 0 | 0 | none |
|
||||
| no_input | 1381 | 0 | 0 | none |
|
||||
| no_input | 1511 | 0 | 0 | none |
|
||||
| no_input | 82312 | 0 | 0 | none |
|
||||
| no_input | 133666 | 0 | 0 | none |
|
||||
|
||||
## Verification and boundaries
|
||||
|
||||
Seven unit tests; all five real motifs CPU/GPU event, count, final V/g/refractory, queue and timestep history exact; five archived deterministic-v2 full-brain trials exact.
|
||||
Every selected edge_id, source CSR row, destination and original weight validated; generated prepared inputs hashed. All original snapshot files and Phase 4AA/4AB model/experiment manifests checked before and after.
|
||||
Diagnostic wall time including compilation and observation: 95.398 s. Peak allocated process VRAM: 285994496 bytes. Not a runtime-speed benchmark.
|
||||
Original author code was inspected, not executed; no Brian2 oracle run and no new package installation. ComfyUI was not stopped or unloaded.
|
||||
|
||||
## Decision
|
||||
|
||||
The usable-learning gate remains unmet: independent evaluation, A-vs-neutral-B specificity in MBON/downstream, multiple training seeds and conditioning controls still required.
|
||||
Before any full-network alternative, obtain explicit separate model authorization and distinguish each semantic component; do not claim g reset alone reproduces the author implementation.
|
||||
No evidence here permits changing eta, thresholds, gain, frozen APL calibration, transmitter signs or the plastic mask.
|
||||
|
||||
## Primary provenance
|
||||
|
||||
[Shiu et al. 2024](https://www.nature.com/articles/s41586-024-07763-9)
|
||||
[Pinned author implementation](https://github.com/philshiu/Drosophila_brain_model/blob/91bdd1e7dcf193f3e7ca5a8933497fcef63b7960/model.py)
|
||||
Downloaded source, MIT license, environment specification, exact commit, SHA-256 and source clauses retained under external-resources and audit.
|
||||
Reference in New Issue
Block a user