Skip to content

model.md

Latest commit

 

History

History
115 lines (95 loc) · 5.52 KB

model.md

File metadata and controls

115 lines (95 loc) · 5.52 KB
Raw

The Fly Who Does QA - Model

Status: EXPERIMENTAL / UNVALIDATED. This document describes what the model actually does, so results are reproducible -- it is not a claim that any of these choices are biologically validated or that the resulting verdicts are accurate. Follow the DoomFly convention: report negative results honestly.

Ground truth about the connectome

Reused unmodified from MaleCNS v1.0 via neuprint (neuprint-cns.janelia.org, dataset male-cns:v1.0), fetched with scripts/fetch_connectome.py (neuprint.fetch_traced_adjacencies, traced non-cropped neurons only). Confirmed cell populations used by this project (counts as of the fetch that produced manifest.json in the export directory -- always check that file for the actual numbers you're running against, these will drift slightly as proofreading continues):

Role Type(s) matched Count in the built graph
R1-R6 input R1-R6 (exact) 1,394
R8 input R8_unclear, R8d, R8p, R8y, R7R8_unclear 1,414
Defect-score output DNp20 (L/R pair) 2 (bodyId 10059=R, 10162=L)
Confidence output DNpe017 2 (bodyId 10527, 555871)
Reward (plasticity, not yet wired) PAM11 15
Aversive (plasticity, not yet wired) PPL101 2

Sign convention for the weight matrix (connectome_data.py)

Real anatomical connectivity is unsigned (synapse counts). Sign is assigned per presynaptic neuron from predictedNt (neuprint's per-neuron predicted neurotransmitter field):

Predicted transmitter Sign
acetylcholine + (excitatory)
GABA − (inhibitory)
glutamate − (inhibitory)
histamine − (inhibitory)
dopamine, octopamine, serotonin, unclear, anything else + (excitatory)

This is a documented simplification, not a claim about ground-truth physiology (glutamate and histamine, in particular, have context-dependent effects in the insect CNS; neuromodulators like dopamine are not simple fast excitation). It is a discrete choice a reader can question and change.

Each postsynaptic neuron's incoming row is then normalized so sum(|weight|) == 1 -- this keeps the rate model numerically stable across neurons with wildly different raw synapse-count totals; it does not change which neurons connect to which, only the relative magnitude used in the simulation.

Encoder (encoder.py) -- an invented proxy, not a validated visual model

There is no retinotopic mapping. Given a precheck-passed image (alpha already flattened onto neutral gray, see precheck.py):

  1. Letterbox-resize to a configurable grid (--grid-size, default 256x256), preserving aspect ratio, padded with neutral gray (128,128,128) -- no cropping, no distortion.
  2. For R1-R6: resize the letterboxed image to an R x C grid where R*C >= len(r1_r6_ids) (near-square, C = ceil(sqrt(n)), R = ceil(n/C)), convert to grayscale, flatten row-major. Each of the sorted R1-R6 body IDs is assigned exactly one grid cell's luminance (0..1) in that flattened order.
  3. For R8: same grid-and-flatten procedure, but the hue channel (HSV) of the letterboxed image is used as the chrominance summary.
  4. Injected current = the grid value directly (linear, gain=1.0 by default) -- there is no spike encoding here, values are fed as constant external drive for every simulation step (static-frame mode).

Sorting body IDs before assignment makes the mapping deterministic and reproducible across runs of the same connectome export; it is arbitrary with respect to actual retinal topology.

Simulator (simulator.py) -- discrete-time leaky rate model, not spiking

r <- (1 - leak) * r + leak * relu(W @ r + I_ext)

run for a fixed number of --steps (default 30) with I_ext held constant throughout (static-frame mode: "one image = one frame held for N steps"). This is explicitly NOT a spiking or biophysical (Hodgkin-Huxley / LIF) model -- it is the cheapest dynamical system that (a) uses the real signed, weighted connectivity, (b) can settle to a steady state, and (c) is fast enough to run per-image. Tiled-scan mode (Section 1, optional) would run this same loop once per tile with carried-over state between tiles, and has not yet been implemented.

Decoder (decoder.py)

defect_score = rate(DNp20_right) - rate(DNp20_left), confidence_signal = sum(rate(n) for n in DNpe017). The verdict is decided from confidence_signal (low -> fail), not defect_score -- this is a deviation from the spec's literal DoomFly-derived mapping, made after direct evidence it was actively wrong for this encoder. defect_score is still computed and shown, for transparency and future work, but does not drive the verdict. Thresholds (DecoderThresholds) are placeholders (calibrated=False) until a calibration script has run against a labeled set -- every log line and the CLI banner say so explicitly while calibrated is false.

Known limitations (say the negative results, don't hide them)

  • No spiking dynamics, no realistic synaptic time constants, no refractory period.
  • Sign-by-predicted-neurotransmitter is a simplification; confidence in those predictions varies per neuron (predictedNtConfidence in the raw export is available but not currently used to weight or filter).
  • The encoder's grid mapping has no relationship to real ommatidia positions or the optic lobe's actual retinotopic map.
  • Photoreceptor and descending-neuron identities are taken from MaleCNS type labels as-is; no manual verification of individual body IDs beyond checking that the expected DNp20/DNpe017 counts (2 each) hold.