Research map / Architectures and efficiency / CNN design
Understanding Image Representations by Measuring Their Equivariance and Equivalence
Karel Lenc, Andrea Vedaldi
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 |
|---|---|---|---|
| add vertical-flip equivariant map at Conv1 instead of leaving features unchanged | AlexNet · ILSVRC12 val (vertical flip) | top-1 error: 0.75 -> 0.43 | Improved |
| add rotation 90 equivariant map at Conv5 | AlexNet · ILSVRC12 val (Rot 90) | top-1 error: 0.75 -> 0.53 | Improved |
| learn equivariant mapping for horizontal flip | AlexNet · ILSVRC12 val (horizontal flip) | top-1 error: 0.44 -> 0.44 | No clear effect |
| learn equivariant mapping for scaling by 2^-1/2 at Conv5 | AlexNet · ILSVRC12 val (scale) | top-1 error: 0.61 -> 0.50 | Improved |
| use task-oriented loss instead of feature-reconstruction forward selection from Conv3 onwards (ablation) | CNN equivariant map · ILSVRC12 val (vertical flip) | top-1 error | Improved |
| use forward selection instead of task-oriented loss up to Conv2 (ablation) | CNN equivariant map · ILSVRC12 val (vertical flip) | training samples needed | Improved |
| use sparse regression (forward selection) instead of least squares (ablation) | HOG equivariant map · HOG 5x5 cell array | Hellinger reconstruction error | Improved |
| add structured sparsity with neighbourhood m=1 instead of dense (m=inf) (ablation) | forward selection k=5 · HOG regression cost 9x9 | learning time (s): 281.18 -> 5.91 | Improved |
| add structured sparsity with neighbourhood m=3 instead of dense (m=inf) (ablation) | forward selection k=5 · HOG regression cost 5x5 | learning time (s): 12.21 -> 3.90 | Improved |
| stitch Places-trained first layers (Conv5) instead of ImageNet-trained into AlexNet | Franken-CNN · ILSVRC12 val | top-1 error: 0.50 -> 0.65 | Hurt |
| use a learned stitching layer instead of identity map (ablation) | Franken-CNN · ILSVRC12 val | top-1 error: >99% -> 0.43 | Improved |
| use equivariant regression with M_g at Conv5 instead of direct regression | cat head rotation regressor · VOC07 cat head rotation | rotation error (degrees): 10.1 -> 13.4 | Hurt |
Closest papers by meaning
- The Lie Derivative for Measuring Learned Equivariance (ICLR 2023 · similarity 0.90)
- Empowering Networks With Scale and Rotation Equivariance Using A Similarity Convolution (ICLR 2023 · similarity 0.90)
- Discriminating image representations with principal distortions (ICLR 2025 · similarity 0.88)
- On Making $SE(2)$-Invariant Networks Optimal (NeurIPS 2026 · similarity 0.88)
- Interpreting Equivariant Representations (ICML 2024 · similarity 0.88)
- Probing Equivariance and Symmetry Breaking in Convolutional Networks (NeurIPS 2025 · similarity 0.87)
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.
