Research map / Architectures and efficiency / Pruning and efficient tuning
Theoretical Guarantees for One-Shot Magnitude Pruning and Compute-Adaptive Early Exit
Erdem Koyuncu
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 |
|---|---|---|---|
| move the exit head to a deeper layer (higher exit-to-final correlation) (ablation) | frozen-backbone early exit · SST-2 (DistilBERT) | log-log slope of excess error vs compute gap: 1.69 -> 11.75 (layer 1 to layer 5) | Improved |
| move the exit head to a deeper layer (ablation) | frozen ViT-Base early exit · CIFAR-100 | log-log slope of excess error vs compute gap: 1.56 -> 8.78 (layer 3 to layer 9) | Improved |
| remove weight renormalization after pruning (ReLU network) (ablation) | one-shot magnitude pruning · CIFAR-10 CNNs | output distortion | Hurt |
| use sign activation instead of ReLU in the pruned network (ablation) | VGGNet7 pruning · CIFAR-10 | accuracy under sparsity | Hurt |
| use MobileNetV2 (ReLU6, 53 layers) instead of 7-layer ReLU VGGNet (ablation) | one-shot magnitude pruning · CIFAR-10 | accuracy under sparsity | Hurt |
| early exit on multi-class ViT: theory exponent vs empirical slope | Theorem 5.2 predictions · CIFAR-100 | agreement of slope with theory: theory 0.53, 0.60, 0.77 vs empirical 1.56, 3.16, 8.78 | Mixed |
Closest papers by meaning
- Pruning Neural Networks at Initialization: Why Are We Missing the Mark? (ICLR 2021 · similarity 0.89)
- An Operator Theoretic View On Pruning Deep Neural Networks (ICLR 2022 · similarity 0.88)
- Pruning neural networks without any data by iteratively conserving synaptic flow (NeurIPS 2020 · similarity 0.88)
- On the Predictability of Pruning Across Scales (ICML 2021 · similarity 0.87)
- Robust Pruning at Initialization (ICLR 2021 · similarity 0.87)
- How Sparse Can We Prune A Deep Network: A Fundamental Limit Perspective (NeurIPS 2024 · similarity 0.86)
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.
