Skip to content

F2 — Connectome-native navigation: per-cell-type parameter fitting (compass → FB → PFL3 → DNa02) #7

Description

@kkokosa

The fly's own navigation circuit is in the checkpoint and silent in the uniform-LIF model (measured: under wind JO-C/E and odor ORN_DM1 input, DNa02 0–8 Hz, EPG 3–13 Hz symmetric, PFL3 / hΔC ≈ 0 — PLAN.md §12 finding p). Real flies steer through wind (WPN → LNa → PFNa; Suver 2019, Okubo 2020) and odor meeting in the fan-shaped body (hΔC; Matheson 2022) against the EPG compass ring attractor (Seelig & Jayaraman 2015; Kim 2017; Turner-Evans 2017/2020; Hulse 2021), with PFL3 turning goal − heading into a DNa02 steering command (Westeinde 2024; Mussells Pires 2024). Today that decision is our decoder (surge / tumble / cast).

Route, following Lappalainen et al. 2024 (connectome-constrained model of the visual system: synapses fixed from the connectome, a few parameters per cell type fitted on a task, physiology predicted).

Deliverables

  1. Experiments first: recover the EPG / PEN / Δ7 ring order from the connectivity (circular ordering of the EPG↔PEN subgraph); inject a compass bump; test persistence, single-ness (Δ7), rotation under PEN drive, and whether PFNa wind + FB odor reach PFL3 / DNa02. Expected: the bump dies — that failure is the objective.
  2. Per-cell-type parameters (synaptic gain, τm, tonic current, threshold): type-gain tables in the .dfb builder + SimulationOptions.TypeGains.
  3. Subgraph checkpoint extractor (CX + wind + olfactory + steering, ~5–10 k neurons) so the fast backend evaluates hundreds× real time.
  4. Fitting: CMA-ES / evolution strategies on the spiking model against quantitative targets — EPG bump persistence and rotation, hΔC & PFN wind tuning, PFL3 L−R ∝ sin(goal − heading), DNa02 asymmetry — or a differentiable rate surrogate trained by backprop and ported back.
  5. Validation on behaviour never fitted: ON surge / OFF cast (Álvarez-Salvado 2018; van Breugel & Dickinson 2014) in the room demo with the surge/cast decoder switched off and yaw read from DNa02 L−R.

Acceptance: only the encoders and the last motor mapping remain as adapters. Risk: the LIF with per-type gains may still not hold a bump → a neuron-model term (adaptation, conductance synapses) is next — still the connectome, with better dynamics.

Context: roadmap, PLAN.md §11 (F2) and §12 finding p.

Activity

Sign up for free to join this conversation on GitHub. Already have an account? Sign in to comment

Metadata

Metadata

Assignees

No one assigned

    Labels

    featureA roadmap feature (PLAN.md §11)run-the-real-flyMoves the sensory/decision boundary from adapters into the connectome

    Projects

    No projects

      Milestone

      No milestone

      Relationships

      None yet

      Development

      No branches or pull requests

      Issue actions