BACK TO INDEX
CASE-2026-016 // FRUITFLY
RESTRICTEDComputational Neuroscience / Simulation

FruitFly

STATUS:ACTIVE
RISK:HIGH

EXECUTIVE SUMMARY

FruitFly is a whole-brain simulation of the male fruit fly nervous system built on published 2024 connectome parameters, extended with an image-driven visual environment and an output-neuron readout layer. The project simulates the full MaleCNS v1.0 network with LIF neuron dynamics, reconstructs the LPLC2 → DNp01 escape-reflex circuit with an R7 control group, and then closes the loop: synthetic images are fed in, motor-neuron output is mapped to actions, and a cursor is driven through a scroll-and-select environment. The purpose is to measure what a real connectome can actually encode, and to document honestly where it cannot.

PROBLEM STATEMENT

Simulated neural networks typically demonstrate capability by construction rather than by measurement, so there is no ground truth for whether a biological architecture solves a task or merely carries enough parameters to appear to. To make the simulation scientifically meaningful, the network must be instantiated from published parameters, the target circuit must be verifiably isolated, and the readout must be measured against a predefined separability criterion with negative results reported as honestly as positive ones.

INVESTIGATIVE FINDINGS

  • ▸Signed weight caches must be rebuilt whenever weight construction changes, otherwise stale sign conventions silently corrupt results
  • ▸The R7 control must remain correctly silent, which validated the sign fix that anchored the circuit
  • ▸LPLC2 sits at a refractory ceiling of 240 spikes under the pool-wide drive protocol, so it reports drive rather than image identity
  • ▸Recurrent amplification inside the LC pool self-sustains at every input rate, so input-rate scaling cannot control saturation
  • ▸Photoreceptors are uniformly histamine sign -1, so Poisson drive hyperpolarises and silences the pathway rather than exciting it
  • ▸Temporal integration cannot rescue separability when every output neuron fires an identical refractory-clipped count per bin

BEHAVIORAL ANALYSIS

The project is built around the discipline of refusing to claim success. When a separation criterion fails, the failure is characterised mechanistically and reported rather than reframed. The final opponent-drive result is presented exactly as measured: the pathway propagates contrast presence and normalises away pattern identity, with flat images yielding zero drive and zero response. Reporting that limitation is treated as the deliverable, not a failure of it.

TECHNICAL EXECUTION

A signed-weight LIF simulator runs the whole-brain MaleCNS v1.0 network from Shiu et al. 2024 parameters, with a strict cache invalidation rule: changing weight construction requires deleting data/connectome-signed-*.npz/.npy. The escape-reflex path is isolated to LPLC2 → DNp01 with R7 as a negative control that correctly stays silent. A visual subset of 19,267 neurons was extracted and made runnable standalone, and an image interface supplies both dense and structured sparse drive. Twelve steps of measurement followed: a de-saturation sweep proving input-rate cannot control LC saturation; a 40-downstream-neuron readout layer; a deterministic 12-image visual environment with a cursor/scroll/select state machine; a stochastic-response study across 20 seeds that overturned an earlier "21/21 distinct" result as seed-noise-dominated; temporal SNR testing that froze the separation ratio at 0.336 across every window from 0.3 s to 2.0 s; onset-readout analysis showing all signal and all noise live in the first 10 ms; a photoreceptor-stage drive attempt that left the pathway silent at every rate; and finally an ON/OFF opponent contrast-coded drive that propagated successfully. Every layer is deterministic and self-checking, with `--test` gates and a closed-loop replay test proving identical trajectories on repeat runs.

SYSTEM ARCHITECTURE

Image Input Interface (Dense + Structured Sparse Modes) → Drive Generation (ON/OFF Opponent Contrast Coding) → Whole-Brain LIF Simulation (MaleCNS v1.0, Shiu et al. 2024, Signed Weight Cache) → LPLC2 → DNp01 Escape-Reflex Circuit with R7 Silent Control → 40-Downstream-Neuron Neural Readout (Body ID / Type / Soma Side / Spikes / Rate Hz / Normalised Activity, plus generic action channels) → Action Policy Mapping (Left / Right / Select with threshold, margin, and cooldown) → Visual Environment State Machine (NO_OP / LEFT / RIGHT / SCROLL_UP / SCROLL_DOWN / SELECT) → Episode Observation Logs (JSON + CSV).

OPERATIONAL FEATURES

  • ✓Whole-brain MaleCNS v1.0 LIF simulation from published 2024 connectome parameters
  • ✓Isolated LPLC2 → DNp01 escape-reflex circuit with a verifiably silent R7 control
  • ✓19,267-neuron visual/action subset extracted and runnable standalone
  • ✓40-downstream-neuron readout with generic left/right action channels and salience reporting
  • ✓Deterministic 12-image visual environment with a cursor/scroll/select state machine
  • ✓Stochastic-response testing across 20 seeds with explicit d-prime measurement
  • ✓Temporal SNR and onset-readout analysis isolating the mechanism of any separation failure
  • ✓ON/OFF opponent contrast-coded drive as a documented experimental interface
  • ✓Closed-loop mode with replay-determinism verification
  • ✓CSV and JSON observation logging for every run

TECHNOLOGICAL ASSETS

PythonNumPySciPyLIF Neuron ModelConnectome Graph AnalysisConnectomics (Shiu et al. 2024 MaleCNS v1.0)

CURRENT STATUS

Active research build. Twelve measurement steps complete, each with a documented result and a defined next lever. The opponent-coded drive stage propagates but does not separate active images, so the next decision is temporal feature extraction on opponent responses or a deliberate readout redesign. All simulation layers remain deterministic and pass their self-checks.

CONCLUSION

FruitFly demonstrates a defensible method for evaluating whether a biologically grounded neural architecture performs a task. The contribution is not a working vision system but a measured account of where a real connectome succeeds, where it saturates, and which specific mechanisms — recurrent amplification, refractory clipping, uniform photoreceptor polarity — cause each failure. Every claim in the project is traceable to a reproducible run.

EVIDENCE LINKS

REPOSITORY

CASE METADATA

OPENED:2026-09-18
ID:CF-2026-016

TAGS

Computational NeuroscienceNeural SimulationLIF ModelConnectomePythonNumPyScientific ComputingEscape ReflexStochastic Analysis
SESSION ACTIVELAST SYNC: --:-- --/TRACE:HOME/CASES/CF-2026-016//ACTIVE[ESC] to exit contextreturning may invalidate findings
○ STANDARDCLEARANCE: LEVEL 2