Fig 5. LN and CNN exemplar model performance.

A. Scatter plot comparing prediction correlation for exemplar pop-LN and 1Dx2-CNN models for each neuron in the A1 dataset. The 1Dx2-CNN model had significantly higher prediction correlation for 376 of the 777 neurons. B. Scatter plot comparing the 2D-CNN and 1Dx2-CNN models, plotted as in A. Prediction correlations were comparable in this case, but the 1Dx2-CNN model still represents a small overall improvement (signed-rank test, p = 1.10 x 10⁻⁸). C. Median prediction correlation in A1 for exemplar models representing each architecture. All differences were statistically significant (signed-rank test, 1D-CNN vs. 2D-CNN: p = 9.21 x 10⁻³, other comparisons: p < 10⁻⁷). D. Prediction correlation for PEG data, plotted as in C. Again, the pop-LN model had the lowest prediction correlation (median 0.46). The difference between 1Dx2-CNN and 2D-CNN was not significant (0.58 vs. 0.57, respectively, signed-rank test, p = 0.883), but all other differences were (p < 10⁻⁴). Although overall prediction correlation was lower for PEG than A1, the relative difference in performance between models was the same for both areas, and the 1Dx2-CNN model was the best-performing in both areas.