Up to eight class examples lowered mean acoustic prediction error at fixed panel sizes in human iEEG and fMRI

We tested acoustic models of human auditory responses. We asked if prediction improves when fitting includes the test sound class. Separate public studies supplied signed iEEG means and BOLD fMRI. Development used four people from each study. We fixed the code before tests in 20 other patients and 16 other adults. For each test class, one fitting panel contained only other classes. A second panel replaced at most eight acoustic matches with examples from the test class. Panel sizes and test sounds were equal. The encoding models used only acoustic features and fixed linear and radial basis kernels. The iEEG mean benefit was 5.175 [3.275, 7.075] points. The BOLD benefit was 2.572 [2.404, 2.741] points. These are percentage points of error from a common mean prediction. Brackets give 95% person Student limits. Positive counts were 18/20 and 16/16. With the test class absent, mean prediction skills were 5.766 and 33.808 points. Both shared claims passed Holm correction, p=0.00027666. A later eight-person MRI check gave mean benefit 2.426 points. An added diagnostic was exploratory. It placed 89.6% and 96.9% of the benefits in the mean residual term. This term contains finite test noise. Residual products from different sounds gave 87.4% and 97.6%. That control omitted three small MRI classes. Some class results were uncertain. Early electrical skill and skill with weaker regularization were also uncertain. A simulation with continuous acoustic responses also gave a positive benefit. The result applies to this fixed panel replacement and prediction rule. It does not isolate semantic information or a common neural cause. Cohort, processing, feature, class and coverage differences limit scope. The package gives code, permitted electrical inputs, source hashes, all results, figures and repeat steps.

human neuroscience · intracranial EEG · fMRI · auditory cortex · acoustic prediction · sound classes · generalization · matched training panels