Research map / Label-efficient and robust learning / Recognition and long-tail
In What Ways Are Deep Neural Networks Invariant and How Should We Measure This?
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 |
|---|---|---|---|
| Train a standard CNN with rotation augmentation instead of random weights | CNN · Out-of-distribution datasets under C8 rotations | latent-space C8-EED | Helped |
| Use a C8-equivariant architecture instead of a standard CNN | CNN trained with rotation augmentation · Rotated MNIST | channelwise C8-EED | Helped |
| Use a C8-equivariant architecture instead of a standard CNN | CNN trained with rotation augmentation · Rotated MNIST | Euclidean latent-space C8-EED | Helped |
| Use a C8-equivariant architecture instead of a standard CNN | CNN trained with rotation augmentation · Rotated MNIST | cosine-similarity latent-space C8-EED | Hurt |
| Add rotation augmentation during training (ablation) | Standard CNN · Rotated MNIST | channelwise C8-EED across layers | No clear effect |
| Add rotation augmentation during training (ablation) | ResNet50, AlexNet, and DenseNet121 · xView Maritime | latent-space C8-EED | Hurt |
| Initialize with ImageNet pretrained weights instead of random weights | ResNet50, AlexNet, DenseNet121, and LeViT192 · xView Maritime | Euclidean latent-space C8-EED | Hurt |
| Use contrastive SimCLR training with reflection augmentations | ResNet50 · ImageNet test with vertical reflections | latent-space reflection EED | Helped |
| Use self-supervised DINO training instead of supervised training | ResNet50 · ImageNet test with vertical reflections | latent-space reflection EED | Mixed |
| Measure latent invariance using cosine similarity rather than Euclidean distance | CNN and C8-CNN · Rotated MNIST | latent-space C8-EED | Mixed |
| Use random weights instead of trained weights | CNN · Out-of-distribution datasets under C8 rotations | softmax C8-EED | Helped |
| Train a standard CNN with augmentation rather than use random weights | CNN · Out-of-distribution datasets under C8 rotations | softmax C8-EED | Hurt |
Closest papers by meaning
- Learning Invariances in Neural Networks from Training Data (NeurIPS 2020 · similarity 0.91)
- Exploiting Invariance in Training Deep Neural Networks (AAAI 2022 · similarity 0.90)
- The Lie Derivative for Measuring Learned Equivariance (ICLR 2023 · similarity 0.89)
- On Making $SE(2)$-Invariant Networks Optimal (NeurIPS 2026 · similarity 0.88)
- A Probabilistic Approach to Learning the Degree of Equivariance in Steerable CNNs (ICML 2024 · similarity 0.88)
- Interpreting Equivariant Representations (ICML 2024 · 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.