Research map / Adversarial robustness and security / Certified and empirical robustness
Towards Theoretical Analysis of Transformation Complexity of ReLU DNNs
Read the paperCheck your idea against 108,967 papers
AI-extracted summary. Written by AI from the paper's title only (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 |
|---|---|---|---|
| Use adversarial training | RN-20 · CIFAR-10, transformation complexity across layers | transformation complexity H(Σ): 3.845 -> 2.979; 6.079 -> 4.485; 6.671 -> 6.573 | Helped |
| Use adversarial training | RN-32 · CIFAR-10, transformation complexity across layers | transformation complexity H(Σ): 2.718 -> 2.507; 5.660 -> 4.370; 6.817 -> 6.786 | Helped |
| Use adversarial training | RN-44 · CIFAR-10, transformation complexity across layers | transformation complexity H(Σ): 3.938 -> 1.600; 5.426 -> 3.374; 6.395 -> 6.828 | Mixed |
| Use adversarial training | LeNet · CIFAR-10, transformation complexity across layers | transformation complexity H(Σ): 7.624 -> 5.358; 7.417 -> 1.216; 10.966 -> 10.949 | Helped |
| Add the transformation-complexity loss and increase its weight λ | Residual MLPs and ResNets · CIFAR-10 and Tiny ImageNet | transformation complexity H(Σ) | Helped |
| Add the transformation-complexity loss and increase its weight λ | Residual MLPs and ResNets · CIFAR-10 and Tiny ImageNet | training-testing loss gap | Helped |
| Increase the weight λ of the transformation-complexity loss | Residual MLPs · CIFAR-10 | testing loss | Helped |
| Add the transformation-complexity loss | DNNs · CIFAR-10 and Tiny ImageNet | minimum L2 norm of adversarial perturbations | Helped |
| Reduce transformation complexity | ResNet-20/32 · CIFAR-10, adversarial transferability | transferability | Mixed |
| Use lower-complexity models | DNNs · Knowledge consistency across models | knowledge consistency | Helped |
| Increase transformation complexity at similar activation rates | Gating states in a layer · DNNs with comparable activation rates | entanglement TC(Σ) | Helped |
| Increase task complexity beyond a certain limit | Residual target MLPs · Knowledge-distillation tasks with varying complexity | transformation complexity | Mixed |
Closest papers by meaning
- On the Local Complexity of Linear Regions in Deep ReLU Networks (ICML 2025 · similarity 0.89)
- Understanding Visual Feature Reliance through the Lens of Complexity (NeurIPS 2024 · similarity 0.84)
- The Impact of Geometric Complexity on Neural Collapse in Transfer Learning (NeurIPS 2024 · similarity 0.84)
- The Computational Complexity of Counting Linear Regions in ReLU Neural Networks (NeurIPS 2025 · similarity 0.83)
- On the Hardness of Training Deep Neural Networks Discretely (AAAI 2025 · similarity 0.83)
- Theoretical Analysis of Robust Overfitting for Wide DNNs: An NTK Approach (ICLR 2024 · similarity 0.83)
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.