Research map / Architectures and efficiency / CNN design
Deep k-Means: Re-Training and Parameter Sharing with Harder Cluster Assignments for Compressing Deep Convolutions
Read the paperCheck your idea against 108,967 papers
AI-extracted summary. Written by AI from the paper's title only (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 |
|---|---|---|---|
| Retrain with spectrally relaxed k-means regularization before weight sharing instead of directly clustering pretrained weights (ablation) | Deep k-Means on FreshNet-CNN · FreshNet-CNN, CIFAR-10, CR 16 | top-1 accuracy change (%): -$5.95 -> -$1.30 | Helped |
| Retrain with spectrally relaxed k-means regularization before weight sharing instead of directly clustering pretrained weights (ablation) | Deep k-Means on Wide ResNet · Wide ResNet, CIFAR-10, CR 45 | top-1 accuracy change (%): -$16.02 -> -$1.63 | Helped |
| Use row-wise k-means weight sharing at CR 4 | Deep k-Means on TT-conv-CNN · TT-conv-CNN, CIFAR-10, CR 4 | top-1 accuracy change (%): -$0.04 | No clear effect |
| Use row-wise k-means weight sharing at CR 2 | Deep k-Means on TT-conv-CNN · TT-conv-CNN, CIFAR-10, CR 2 | top-1 accuracy change (%): +$0.05 | Helped |
| Increase compression ratio from 45 to 50 | Deep k-Means on Wide ResNet · Wide ResNet, CIFAR-10 | top-1 accuracy change (%): -$1.63 -> -$4.49 | Hurt |
| Compress GoogLeNet convolutional layers at CR up to 3 | Deep k-Means on GoogLeNet · ImageNet ILSVRC12 validation | top-1 accuracy change (%): -$0.26, -$0.17, -$0.36 | No clear effect |
| Increase GoogLeNet compression ratio from 3 to 4 | Deep k-Means on GoogLeNet · ImageNet ILSVRC12 validation | top-1 accuracy change (%): -$0.36 -> -$1.95 | Hurt |
| Use spectrally relaxed k-means energy metrics to estimate hardware-related energy | Deep k-Means energy metrics · AlexNet | R^2 against hardware-based extrapolation: 0.9931 | Helped |
| Use spectrally relaxed k-means energy metrics to estimate hardware-related energy | Deep k-Means energy metrics · GoogLeNet_v1 | R^2 against hardware-based extrapolation: 0.9675 | Helped |
Closest papers by meaning
- Towards K-means-friendly Spaces: Simultaneous Deep Learning and Clustering (ICML 2017 · similarity 0.87)
- DKM: Differentiable k-Means Clustering Layer for Neural Network Compression (ICLR 2022 · similarity 0.86)
- Exploiting Kernel Sparsity and Entropy for Interpretable CNN Compression (CVPR 2019 · similarity 0.86)
- Forget the Data and Fine-Tuning! Just Fold the Network to Compress (ICLR 2025 · similarity 0.84)
- Unveiling the Unseen: Identifiable Clusters in Trained Depthwise Convolutional Kernels (ICLR 2024 · similarity 0.84)
- SKIM: Any-bit Quantization Pushing The Limits of Post-Training Quantization (ICML 2025 · 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.