Research map / Label-efficient and robust learning / Recognition and long-tail
On the Functional Similarity of Robust and Non-Robust Neural Representations
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 |
|---|---|---|---|
| Stitch robust front network into non-robust end network (accuracy-optimized affine stitching) | ResNet-18 on CIFAR-10 (fR1 front, fN1 end) · CIFAR-10 | plain accuracy | No clear effect |
| Stitch non-robust front into robust end network with accuracy-optimized stitching, robustness lost after first few layers | ResNet-18 on CIFAR-10 (fN1 front, fR1 end) · CIFAR-10 | robust accuracy | Hurt |
| Robust-robust stitching with adversarial stitching layer (alpha=1) in late layers | ResNet-18 on CIFAR-10 (fR1 front, fR1 end) · CIFAR-10 | robust accuracy | Helped |
| Plain-robust adversarial stitching: alpha=0.5 vs alpha=1 for robust end network in early layers (ablation) | Plain front, robust end stitching · CIFAR-10 | plain accuracy | Helped |
| Robust-plain adversarial stitching: alpha=0.5 instead of alpha=1 in later layers (ablation) | Robust front, plain end stitching · CIFAR-10 | plain accuracy | Helped |
| Plain-plain adversarial stitching with alpha=1 (100% adversarial examples) (ablation) | Non-robust front and end networks · CIFAR-10 | plain accuracy | Hurt |
| Cross-task stitching in last layers (CIFAR-10 and SVHN networks) | Last layers of ResNet-18 networks · CIFAR-10 / SVHN cross-task | plain accuracy | No clear effect |
| Stitching robust front into robust end self-stitching with accuracy objective | Robust ResNet-18 self-stitching · CIFAR-10 | robust accuracy | Hurt |
Closest papers by meaning
- Understanding Robust Learning through the Lens of Representation Similarities (NeurIPS 2022 · similarity 0.91)
- Adversarially Robust Few-shot Learning via Parameter Co-distillation of Similarity and Class Concept Learners (CVPR 2024 · similarity 0.88)
- Revisiting Model Stitching to Compare Neural Representations (NeurIPS 2021 · similarity 0.88)
- Functional Alignment Can Mislead: Examining Model Stitching (ICML 2025 · similarity 0.88)
- How Not to Stitch Representations to Measure Similarity: Task Loss Matching Versus Direct Matching (AAAI 2025 · similarity 0.87)
- Feature Averaging: An Implicit Bias of Gradient Descent Leading to Non-Robustness in Neural Networks (ICLR 2025 · similarity 0.86)
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.