Table 2 Comparisons of performance for the activation functions Beaf, ELU, and GeLU on PCBA, BACE,Tox21 and SIDER datasets (higher values are better)
From: Double-head transformer neural network for molecular property prediction
| Â | GeLU | ELU | Beaf |
|---|---|---|---|
PCBA | \(0.806\pm 0.002\) | \(0.663\pm 0.006\) | \(0.821\pm 0.005\) |
BACE | \(0.928\pm 0.019\) | \(0.909\pm 0.022\) | \(0.923\pm 0.035\) |
Tox21 | \(0.843\pm 0.025\) | \(0.840\pm 0.049\) | \(0.847\pm 0.015\) |
SIDER | \(0.652\pm 0.027\) | \(0.628\pm 0.012\) | \(0.679\pm 0.015\) |