Research map / Architectures and efficiency / CNN design
SPIN: An Empirical Evaluation on Sharing Parameters of Isotropic Networks
Chien-Yu Lin, Anish Prabhu, Thomas Merth, Sachin Mehta, Anurag Ranjan, Maxwell Horton, Mohammad Rastegari
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 |
|---|---|---|---|
| weight fusion init from pretrained net (channel weighted mean) instead of random init (ablation) | WS-ConvMixer-768/32 share rate 2 · ImageNet-1k | top-1 accuracy (%): 73.23 -> 75.15 | Improved |
| fuse weights without pretrained initialization (Mean) (ablation) | WS-ConvMixer · ImageNet-1k | top-1 accuracy (%): 74.29 -> 74.25 | No clear effect |
| initialize with pretrained weights when using Mean fusion (ablation) | WS-ConvMixer · ImageNet-1k | top-1 accuracy (%): 74.25 -> 75.28 | Improved |
| add lightweight affine weight transformation (G=64) (ablation) | WS-ConvMixer share rate 2 · ImageNet-1k | top-1 accuracy (%): 73.29 -> 73.36 | No clear effect |
| add affine weight transformation with G=1 (ablation) | WS-ConvMixer · ImageNet-1k | top-1 accuracy (%): 73.29 -> 72.87 | Hurt |
| sequential instead of strided sharing mapping (ablation) | WS-ConvMixer uniform · ImageNet-1k | top-1 accuracy (%): 72.80 -> 73.29 | Improved |
| random sharing mapping instead of sequential (ablation) | WS-ConvMixer uniform · ImageNet-1k | top-1 accuracy (%): 73.29 -> Diverged | Hurt |
| share front layers instead of back layers (ablation) | WS-ConvMixer sequential · ImageNet-1k | top-1 accuracy (%): 72.35 -> 73.31 | Improved |
| share only pointwise conv rather than also BN, bias, depthwise (ablation) | WS-ConvMixer share rate 2 · ImageNet-1k | top-1 accuracy (%): 73.17 -> 73.29 | Improved |
| share BatchNorm as well as other components (ablation) | WS-ConvMixer · ImageNet-1k | top-1 accuracy (%): 73.17 -> Diverged | Hurt |
| weight sharing (rate 8) instead of depth scaling at equal cost | ConvMixer · ImageNet-1k | top-1 accuracy (%): 57.75 -> 67.19 | Improved |
| weight sharing rate 2 with fusion on ConvMixer-1536/20 | ConvMixer-1536/20 · ImageNet-1k | top-1 accuracy (%): 78.03 -> 78.47 | Improved |
Closest papers by meaning
- Weight Fixing Networks (ECCV 2022 · similarity 0.85)
- AlphaNet: Improved Training of Supernets with Alpha-Divergence (ICML 2021 · similarity 0.85)
- Do We Need Anisotropic Graph Neural Networks? (ICLR 2022 · similarity 0.85)
- WeightNet: Revisiting the Design Space of Weight Networks (ECCV 2020 · similarity 0.85)
- Role of Locality and Weight Sharing in Image-Based Tasks: A Sample Complexity Separation between CNNs, LCNs, and FCNs (ICLR 2024 · similarity 0.84)
- Training a Predictive Coding Network on ImageNet using Equilibrium Propagation (NeurIPS 2026 · similarity 0.83)
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.
