Figure 6.

Overfitting by the FIR model. We manipulated two characteristics of the POLY model (Table I) and evaluated the effect on fit accuracy (gray line) and prediction accuracy (black line). For these graphs we selected voxels with a minimum SNR of 10 under the POLY model (n = 1,730). Dots indicate the median across voxels, and error bars indicate ±1 SE (bootstrap procedure). (A) The effect of HDR window duration on fit accuracy and prediction accuracy. The x‐axis indicates the HDR window duration used in the model; the y‐axis indicates explained variance. Prediction accuracy was maximized at a duration of 9 s. This indicates that, on average, estimating HDRs beyond 9 s resulted in overfitting and reduced model generalizability. (B) The effect of the number of event types on fit accuracy and prediction accuracy. Based on SNR estimates obtained under the POLY model, we refit the model including only the top event types with respect to SNR. (Because different voxels respond to different event types, the included event types varied on a voxel‐by‐voxel basis.) The x‐axis indicates the number of event types; the y‐axis indicates explained variance. Prediction accuracy was maximized at three event types. This indicates that, on average, estimating more than three event types resulted in overfitting and reduced model generalizability. This result is explained by the fact that voxels in visual cortex are often highly selective for spatial position, in such a way that stimuli positioned at nonpreferred locations produce no discernable activation.