Research map / Architectures and efficiency / RNNs and normalization
KLAS: Using Similarity to Stitch Neural Networks for an Improved Accuracy-Efficiency Tradeoff
Debopam Sanyal, Anantharaman S. Iyer, Alind Khare, Trisha Jain, Akshay Jajoo, Myungjin Lee, James Clayton Kerce, Alexey Tumanov
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 KL-divergence-based stitch selection (KLAS) instead of heuristic selection | SN-Net · ImageNet-1K, Swin | AUC: 0.8345 -> 0.8950 | Improved |
| select stitches with KLAS on ImageNet-1K | SN-Net · ImageNet-1K | top-1 accuracy / FLOPs: up to 1.21% higher accuracy, or 1.33x fewer FLOPs | Improved |
| use far stitching (Swin Ti-B) instead of nearest stitching (Ti-S-B) (ablation) | Swin stitching · ImageNet-1K | accuracy at equal FLOPs: up to 0.9% | Improved |
| use far stitching (Ti-B) instead of nearest stitching for DeiT (ablation) | DeiT stitching · ImageNet-1K | accuracy-efficiency curve | Hurt |
| use KLAS block selection with SN-Net's anchors | SN-Net · ImageNet-1K, Swin and DeiT | AUC | Improved |
| use naive Min-KL selection instead of target-aware stitch score (ablation) | KLAS · DeiT and Swin | accuracy-efficiency curve | Hurt |
| use KLAS on dense prediction | SN-Net · ADE20K, Mask2Former Swin-T/B | mIoU: up to +0.9% mIoU | Improved |
| use KLAS instead of ESTA layer-count heuristic for LLM stitching | ESTA · TruthfulQA, Llama 3.2 1B to 3B | ROUGE-1 and ROUGE-2: up to 0.017 ROUGE-1 and 0.033 ROUGE-2 | Improved |
| use KLAS stitched models instead of early-exit cascades | Swin-S and Swin-B cascade · ImageNet-1K | top-1 accuracy: over 0.4% above best cascade at more than 12 GFLOPs | Improved |
| use KLAS instead of cascade for ResNet-18 to Swin-B | cascade · ImageNet-1K | top-1 accuracy: 76.13 -> 79.60 at config 1 | Improved |
| train probes with shuffled labels (ablation) | stitch score · Swin | Pearson correlation with accuracy drop: 0.84 -> 0.19 | Hurt |
| use threshold tau of 10% instead of 5% (ablation) | KLAS · Swin, 20 buckets | accuracy and AUC | Hurt |
Closest papers by meaning
- Efficient Stitchable Task Adaptation (CVPR 2024 · similarity 0.86)
- How Not to Stitch Representations to Measure Similarity: Task Loss Matching Versus Direct Matching (AAAI 2025 · similarity 0.86)
- Reconstructing the Image Stitching Pipeline: Integrating Fusion and Rectangling into a Unified Inpainting Model (NeurIPS 2024 · similarity 0.86)
- Stitched Value Model for Diffusion Alignment (NeurIPS 2026 · similarity 0.86)
- Towards Robust Image Stitching: An Adaptive Resistance Learning against Compatible Attacks (AAAI 2024 · similarity 0.85)
- Functional Alignment Can Mislead: Examining Model Stitching (ICML 2025 · similarity 0.85)
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.
