Keeping full cell responses raised mean odor decoder accuracy by 0.96 percentage points over one fitted component in 67 sister groups from three mice

We tested whether cells with a traced common glomerular input can be compressed to one shared response component without loss of odor decoding accuracy. We used 372 mitral or tufted cells in 67 anatomical sister groups from three anesthetized mice. Each feature was the mean calcium signal during the first two seconds after odor onset. A fixed nearest-centroid decoder trained on other repeats and tested one excluded repeat. Both models used the same cells. One kept all responses. The other kept the first principal component of the training odor means. Equal-mouse accuracy was 5.54% with full responses and 4.58% with one component, for panels of 47 or 48 odors. The increase was 0.96 percentage points. A descriptive group resampling range within these mice was 0.49 to 1.43 points. The increases were +2.10, -0.27, and +1.04 points in the three mice. Residual odor shuffles within delivery modules gave lower mean accuracy. A later pre-odor subtraction check kept a mean increase of 0.81 points, but reduced the third mouse gain to almost zero. A stationary Gaussian rank-one reference also gave a smaller mean increase. The source study already showed sister response diversity and odor decoding as cells were added. Our contribution keeps cell count fixed and tests fitted common-component compression. The results show a small loss for this compression and decoder. They do not establish the best possible code dimension, a synaptic mechanism, or a behavioral benefit. Few repeats, three mice, and baseline sensitivity limit the result. The package preserves compact data, source records, code, all predictions and controls, figures, and repeat instructions.

olfactory bulb · sister cells · structure and function · odor decoding · calcium imaging · principal component · anatomical convergence