Research map / Architectures and efficiency / RNNs and normalization
Trainable Highly-expressive Activation Functions
Irit Chelly, Shahaf E. Finder, Shira Ifergane, Oren Freifeld
Read the paperCheck your idea against 137,217 papers
AI-extracted. The measured effects below were read by AI from the paper's full text (no summary yet). Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| replace fixed activations with DiTAC | MLP · 2D-GMM classification | top-1 accuracy (%): 86.8 -> 89.06 (vs Swish) | Improved |
| replace activations with DiTAC | MLP · MNIST | top-1 accuracy (%): 97.99 -> 98.29 (vs GELU) | Improved |
| replace activations with DiTAC | 1D/2D regression MLP · 2D function regression | MSE: 0.008 -> 0.004 (vs PReLU) | Improved |
| replace activations with DiTAC | MobileNet-V3 and ResNet · ImageNet-50 | top-1 accuracy (%): up to 0.6% on MobileNet-V3, 0.12% on ResNet | Improved |
| DiTAC vs LReLU on ResNet-50 | ResNet-50 · ImageNet-50 | top-1 accuracy (%): 93.24 -> 93.24 | No clear effect |
| replace GELU with DiTAC | ConvNeXt-T · ImageNet-1K | top-1 accuracy (%): 82.1 -> 82.4 | Improved |
| replace GELU with DiTAC | Swin-T · ImageNet-1K | top-1 accuracy (%): 81.3 -> 81.5 | Improved |
| replace baseline AF with DiTAC | Swin-T + UperNet · ADE20K | mIoU: 44.4 -> 44.7 | Improved |
| replace baseline AF with DiTAC | ResNet-50 + PSPNet · Cityscapes | mIoU: 77.9 -> 78.4 | Improved |
| replace ReLU with DiTAC in generator and discriminator | DCGAN · CelebA | FID: 133.5 -> 34.6 | Improved |
| replace baseline AF with GELU in GANs | DCGAN · CelebA | FID: 133.5 -> 516.5 | Hurt |
| replace ReLU with DiTAC | BigGAN · CIFAR-10 | FID: 10.3 -> 10.0 | Improved |
Closest papers by meaning
- DiGRAF: Diffeomorphic Graph-Adaptive Activation Function (NeurIPS 2024 · similarity 0.86)
- Entropy-based Activation Function Optimization: A Method on Searching Better Activation Functions (ICLR 2025 · similarity 0.86)
- Efficient Activation Function Optimization through Surrogate Modeling (NeurIPS 2023 · similarity 0.85)
- Gompertz Linear Units: Leveraging Asymmetry for Enhanced Learning Dynamics (NeurIPS 2025 · similarity 0.84)
- Better NTK Conditioning: A Free Lunch from (ReLU) Nonlinear Activation in Wide Neural Networks (NeurIPS 2025 · similarity 0.84)
- Rethinking Neural Nonlinearity as Gating (NeurIPS 2026 · similarity 0.84)
Writing a paper in this area? scime draft is a free LaTeX editor with AI included, real-time co-authors and a check of your references and venue rules before you submit. It also imports OpenAI Prism projects.
