Table 3 Steps of the ablation experiment and corresponding accuracy

Ablation experiment	Cross-scale embedding layer	Dynamic position bias	Short-range attention	Long-range attention	Depth of transformer layers	Accuracy
Complete model	√	√	√	√	√	96.32%
Convolutional layer	Convolutional layer	√	√	√	√	94.38%
Removing dynamic position bias	√	×	√	√	√	93.92%
Only short-range attention	√	√	×	√	√	95.47%
Only long-range attention	√	√	√	×	√	90.78%
Reducing the depth of transformer layers	√	√	√	√	↓	94.53%
Increasing the depth of transformer layers	√	√	√	√	↑	91.10%

Quick links

Search