BANC / BEHAVIORAL FORECASTING
One fly. A shorter horizon.
Can the next position be learned?
Filled white: current · White ring: predicted · Gray square: observed
60% train / 20% validation / 20% test, in time order. Missing-data gaps and split boundaries are excluded.
Fixed ridge model. Training uses only earlier positions and displacements. No neural connections are used in this first experiment.
What this is: an actual trained position predictor using the published behavioral recording associated with the BANC fly. The connectome paper is the research starting point; this model is not a reconstruction of its brain.
Read the paper ↗ · Dataset documentation ↗
RMSE is the square root of the mean squared Euclidean position error. Lower is better. One fly, one recording, overlapping test windows: these are not independent trials or evidence of human forecasting ability. Repeatedly viewing this holdout makes later tuning exploratory.
Data provenance and the next research step
Loading provenance…
Original CSV SHA-256: Loading…
Next step: select a documented circuit from BANC, define neuron dynamics and an observable task, and compare its predictions with this behavioral baseline. Structural connections alone do not specify a working nervous system.