Time-shifted TRF improves prediction accuracy.
<p>The prediction accuracy was measured using Pearson’s correlation, and the significance of improvement was measured using paired Wilcoxon signed rank tests across participants. Electrodes where the prediction improvement is significant (p < 0.05) were labeled with white circles. P values...
Saved in:
| Main Author: | |
|---|---|
| Other Authors: | , , , |
| Published: |
2025
|
| Subjects: | |
| Tags: |
Add Tag
No Tags, Be the first to tag this record!
|
| Summary: | <p>The prediction accuracy was measured using Pearson’s correlation, and the significance of improvement was measured using paired Wilcoxon signed rank tests across participants. Electrodes where the prediction improvement is significant (p < 0.05) were labeled with white circles. P values were corrected using false discovery rate based on Benjamini-Hochberg correction (FDR), with alpha set to 0.05. <b>A,</b> Topographic maps for comparing prediction accuracies of different TRF variants on individual electrodes. The left map (in the black box) shows the improvement of prediction accuracy for a static TRF when adding lexical surprisal as a predictor in addition to the acoustic envelope and onset control predictors. The right three maps then show the additional improvement of three dynamic TRF variants over the static TRF. From left to right, the three variants are amplitude scaling (Amp), time shifting (Time), and their combination (Time&Amp). <b>B,</b> The bar plot quantifies prediction accuracies of the same static TRF (left) and dynamic TRFs (right) averaged across all 128 channels. Text/asterisks in blue in the right panel indicate statistical comparison results between each dynamic TRF and the static TRF. Horizontal connecting bars indicate the comparison between the different dynamic TRF variants. Significance is indicated by * if p < 0.05, ** if p < 0.01, and *** if p < 0.001. <b>C,</b> Convolutional kernel for the amplitude scaling (left) and time-shift (right) parameters. The lag indicates the index relative to the current time step (e.g., 0 and -1 indicate the weights for the current and previous word, respectively). Nat indicates the natural unit of information of the surprisal values. Light blue colored lines indicate the weights of models fitted from individual rounds of cross-validation (each round of which involved training an all but two runs from every participant – i.e., for each participant, two runs were left out for testing). The deep blue colored line indicates the weights averaged across those cross-validation rounds.</p> |
|---|