Research map / Architectures and efficiency / RNNs and normalization
On the Symmetries of Deep Learning Models and their Internal Representations
Charles Godfrey, Davis Brown, Tegan Emerson, Henry Kvinge
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 |
|---|---|---|---|
| increase LeakyReLU negative slope from 0.1 to 1.0 (ablation) | ResNet20 · CIFAR-10 | accuracy (%): 89.4 -> 41.8 | Hurt |
| increase LeakyReLU negative slope from 1e-4 to 0.1 (ablation) | ResNet20 · CIFAR-10 | accuracy (%): 89.3 -> 89.4 | No clear effect |
| increase LeakyReLU negative slope to 0.9 (ablation) | ResNet20 · CIFAR-10 | accuracy (%): 89.4 -> 73.0 | Hurt |
| use orthogonal instead of permutation/scaling invariance in Procrustes | Myrtle CNN · CIFAR-10 layer 3 | Procrustes similarity: 0.6208 -> 0.7724 | Mixed |
| remove ReLU on residual outputs (ablation) | ResNet-50 · ImageNet network dissection | interpretable units under rotation | Mixed |
| replace ResNet-50 with ConvNeXt (no residual activation) (ablation) | ResNet-50 · ImageNet network dissection | interpretable units under rotation | Mixed |
| use constant channel width 512 in Myrtle CNN (ablation) | Myrtle CNN · CIFAR-10 | dissimilarity measures | No clear effect |
| choose alpha by cross validation instead of fixed 0.1 (ablation) | G_ReLU stitching · CIFAR-10 | stitching accuracy | No clear effect |
| l1-regularized stitching lambda 0.001-0.1 vs G_ReLU stitching (ablation) | Myrtle CNN · CIFAR-10 | stitching penalty vs sparsity | Mixed |
| set l1 penalty lambda to 1 (ablation) | l1 stitching · CIFAR-10 | optimization stability | Hurt |
Closest papers by meaning
- Improving Equivariant Networks with Probabilistic Symmetry Breaking (ICLR 2025 · similarity 0.89)
- Parameter symmetries determine representational geometry in overparameterized nonlinear networks (NeurIPS 2026 · similarity 0.88)
- A Tale of Two Symmetries: Exploring the Loss Landscape of Equivariant Models (NeurIPS 2025 · similarity 0.88)
- Symmetry Induces Structure and Constraint of Learning (ICML 2024 · similarity 0.88)
- Symmetries in Overparametrized Neural Networks: A Mean Field View (NeurIPS 2024 · similarity 0.88)
- Learning symmetries via weight-sharing with doubly stochastic tensors (NeurIPS 2024 · similarity 0.88)
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.
