Research map / Label-efficient and robust learning / Recognition and long-tail
Deconfounded Representation Similarity for Comparison of Neural Networks
Tianyu Cui, Yogesh Kumar, Pekka Marttinen, Samuel Kaski
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 |
|---|---|---|---|
| Deconfound CKA (CKA -> dCKA) | CKA · ImageNet-CIFAR pair detection, average over blocks | proportion detected: 0.24, 0.41 -> 0.5, 0.51 | Improved |
| Deconfound RSA (RSA -> dRSA) | RSA · ImageNet-CIFAR pair detection, average over blocks | proportion detected: 0.02, 0.15 -> 0.25, 0.26 | Improved |
| Deconfound CKA in deep blocks | CKA · ImageNet-CIFAR pair detection, blocks 5-8 | proportion detected: 0.0, 0.0 -> 0.0, 0.04 | No clear effect |
| Deconfound CKA across noise levels | CKA · Cross-domain NN identification (CIFAR-10-C) | proportion identified: 0.18 -> 0.29 | Improved |
| Deconfound RSA across noise levels | RSA · Cross-domain NN identification (CIFAR-10-C) | proportion identified: 0.16 -> 0.23 | Improved |
| Deconfound CKA on same-domain inputs | CKA · Same-domain NN identification | proportion identified: 0.65 to 0.8 | Improved |
| Deconfound RSA on same-domain inputs | RSA · Same-domain NN identification | proportion identified: 0.75 to 0.8 | Improved |
| Deconfound CKA for transfer learning | CKA · DomainNet domain similarity correlation | Spearman rho: 0.804 -> 0.895 | Improved |
| Deconfound CKA for transfer learning | CKA · Multi STS-B domain similarity correlation | Spearman rho: 0.358 -> 0.924 | Improved |
| Deconfound CKA for OOD accuracy prediction | CKA · CIFAR-10-C average over corruption levels | Spearman rho: 0.139 -> 0.145 | Improved |
| Deconfound CKA for in-distribution accuracy | CKA · CIFAR-10 ID accuracy | Spearman rho: 0.163 -> 0.167 | No clear effect |
| Add higher-order polynomial terms of input similarity to the regression (ablation) | deconfounding regression · ResNet-18 CIFAR-10 | R^2 and BIC | No clear effect |
Closest papers by meaning
- Differentiable Optimization of Similarity Scores Between Models and Brains (ICLR 2025 · similarity 0.88)
- Spectral Analysis of Representational Similarity with Limited Neurons (NeurIPS 2025 · similarity 0.87)
- Tracing Representation Progression: Analyzing and Enhancing Layer-Wise Similarity (ICLR 2025 · similarity 0.86)
- Not all solutions are created equal: An analytical dissociation of functional and representational similarity in deep linear neural networks (ICML 2025 · similarity 0.86)
- ReSi: A Comprehensive Benchmark for Representational Similarity Measures (ICLR 2025 · similarity 0.85)
- Going Beyond Neural Network Feature Similarity: The Network Feature Complexity and Its Interpretation Using Category Theory (ICLR 2024 · similarity 0.85)
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.
