Research map / Architectures and efficiency / RNNs and normalization
Densely Connected G-invariant Deep Neural Networks with Signed Permutation Representations
Devanshu Agrawal, James Ostrowski
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 |
|---|---|---|---|
| use type 2 signed perm-irreps instead of type 1 | G-DNN · binary multiplication (m=16) | classification accuracy: 50% -> 100% | Improved |
| use type 2 signed perm-irreps instead of type 1 | G-DNN · binary multiplication (m=16) | final binary cross-entropy loss: 0.71 -> 0.00 | Improved |
| replace constrained G-DNN by the larger unraveled architecture (about 6.5x parameters) | G-DNN type 2 · binary multiplication (m=16) | classification accuracy: 100% -> 50% | Hurt |
| initialize unraveled architecture with type 2 initial weights | unraveled architecture · binary multiplication (m=16) | final loss: 11.36 -> 1.33 initial loss; final 0.71 -> 0.70 | No clear effect |
| use G-DNN skip-connected architecture instead of CReLU | architecture space · icosahedral group admissible architectures | number of admissible architectures: 2639 -> 4751 | Improved |
| use G-DNN instead of CReLU at depth 9 | architecture space · icosahedral group | admissible architectures: 0 -> 80 | Improved |
| use G-DNN instead of CReLU at depth 4 | architecture space · group C2^3, |G|=8 | admissible fraction: 238/392 -> 392/392 | Improved |
| use G-DNN instead of CReLU at depth 3 | architecture space · group C2^3, |G|=8 | admissible fraction: 88/434 -> 93/434 | Improved |
| use G-DNN instead of CReLU at depth 4 | architecture space · group C2xC4, |G|=8 | admissible fraction: 34/48 -> 48/48 | Improved |
| use G-DNN instead of CReLU for cyclic C8 and Q8 | architecture space · groups C8 and Q8 | admissible fraction: 9/9 -> 9/9 at depth 2 for Q8 | No clear effect |
| use mixed architecture with type 1 and type 2 signed perm-irreps instead of type 1 or unraveled only | G-DNN · ModelNet40 validation | validation accuracy | Improved |
Closest papers by meaning
- A Classification of $G$-invariant Shallow Neural Networks (NeurIPS 2022 · similarity 0.90)
- The Selective $G$-Bispectrum and its Inversion: Applications to $G$-Invariant Networks (NeurIPS 2024 · similarity 0.87)
- On the VC dimension of deep group convolutional neural networks (NeurIPS 2025 · similarity 0.85)
- Beyond the Permutation Symmetry of Transformers: The Role of Rotation for Model Fusion (ICML 2025 · similarity 0.84)
- MuseGNN: Forming Scalable, Convergent GNN Layers that Minimize a Sampling-Based Energy (ICLR 2025 · similarity 0.84)
- Grokking Group Multiplication with Cosets (ICML 2024 · similarity 0.84)
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.
