Figure 7
Supervised learning evaluation results for the VISp reference data and c-layers: Although the exact numbers are partly sensitive to the quality of the training data, the neuro-biologically inspired feature sets F1–F4 eventually result in labels that more closely resemble the manual ones than those obtained using the image texture features F5 and F6. (A–H) fd=B statistics for the different reference locations (panels) and different feature sets (horizontal axis). The histograms show the mean values over the 60 locations, the error bar the standard deviations, and blue and black dots the maximum and minimum values. Comparison of differences between F1 and F3 or F1 and F4 are, except for M262R3 and M339R6, significant (p < 0.01). In contrast, differences between F3 and F4 are not significant for all cases (p > 0.4). For comparison purposes, the olive lines represent respective statistics when the datasets manually labelled by the second reader are used as the test data. Panels (I,J) contain additional information about the reference c-layers. Panel (I) shows the number of manually discriminated and labelled layers of the reference locations, namely nB. (J) shows the similarity between the refined reference c-layer structures and the refined manually labelled layers structures. (K) The statistics of the self-identical measurement (=re-labelling) of the reference locations, where all the symbols have the same statistical meaning as in panels (A–H). All reference c-layers were selected on the hierarchical level after the 8th clustering iteration. Target c-layers were selected on the last hierarchical level that exhibited a layer number not smaller than the number of reference c-layers.
