Face and house decoding remained after removal of the largest fitted variance axis in selected records from six ECoG patients and five fMRI adults

We tested whether face and house information remains after removal of the largest training variance axis. Separate public human studies supplied ECoG power and BOLD fMRI. One person from each study supplied development. A local plan and code lock preceded tests in six other ECoG patients and five other fMRI adults. Site scales and the removed axis used only earlier data. A centroid reader used the remaining space. ECoG tests used later trials and image IDs absent from training. MRI tests used later whole runs. Mean test AUC was 0.861 and 0.830. Their 95% person Student intervals were [0.791, 0.931] and [0.627, 1.034]. Scores were above chance in 6/6 and 5/5 people. Bounded 95% intervals were [0.307, 1.000] and [0.223, 1.000]. The fixed shared test passed. Its person-and-label conjunction p was 0.03125. All development and control results remain. An exact example gave residual AUC 0.769 with one true category axis and a 10 degree error in the removed axis. Thus residual decoding does not count true neural dimensions. Clinical selection, source masks, different tasks, fixed stimuli, short training sets and measurement filters limit scope. The package supplies code, source hashes and terms, figures, all results and repeat instructions. Original arrays stay external.

human neuroscience · ECoG · fMRI · principal component analysis · category decoding · public data · confirmation