Full decoder transfer gave 75 to 76 percent median efficiency in two adult worm map models at half sharing

We tested transfer of ideal linear source decoders between two adult C. elegans brain maps. The model used positive unit contact effects, independent sites and equal marginal noise. At within-site sharing fraction 0.5, full transfer of signal and covariance gave median test-map efficiency of 0.753 to 0.759 relative to the test optimum. Supplying the test-map signal raised the median to 0.952 to 0.963. The median paired shared-covariance advantage over diagonal training under full transfer was only 0.3 to 0.6 percent. The comparison used 140 common sources and 160 common targets. Lower sharing removed that median advantage. Diagonal regularization did not improve median efficiency in both directions. These are exact conditional model scores, not measured neural information. Two animals, unknown release dependence, unit positive effects, partial coverage and ideal signed readouts limit transfer. Code, aggregate results, source hashes and independent checks support reproduction. External raw data and record-level derivative databases are excluded.

connectomics · C. elegans · decoder transfer · polyadic synapses · noise covariance · linear sensitivity