Research map / Detection and segmentation / CNN features
LoCo: Local Contrastive Representation Learning
Yuwen Xiong, Mengye Ren, Raquel Urtasun
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 |
|---|---|---|---|
| add overlapping shared stage between local units (LoCo) to gradient-isolated local learning (ablation) | SimCLR local learning · ImageNet-1K linear eval, 100 ep | top-1 accuracy: 65.1 -> 66.2 | Improved |
| LoCo instead of GIM on SimCLR | ResNet-50 · ImageNet-1K linear eval | top-1 accuracy: 64.7 -> 69.5 | Improved |
| add one conv block with downsampling to local decoders (ablation) | local contrastive learning · ImageNet-1K linear eval, 100 ep | top-1 accuracy: 60.9 -> 65.1 | Improved |
| add one conv block to end-to-end SimCLR encoder end (ablation) | SimCLR · ImageNet-1K linear eval, 100 ep | top-1 accuracy: 65.7 -> 65.6 | No clear effect |
| remove downsampling from the added decoder conv block (ablation) | local decoder with 1 conv block · ImageNet-1K linear eval, 100 ep | top-1 accuracy: 65.1 -> 63.6 | Hurt |
| deepen local decoder to 2 conv blocks, 1 stage or full network (ablation) | local learning without sharing · ImageNet-1K linear eval, 100 ep | top-1 accuracy: 65.1 -> 65.8 | Improved |
| add two more MLP decoder layers (4 total) (ablation) | local learning without sharing · ImageNet-1K linear eval, 100 ep | top-1 accuracy: 60.9 -> 62.3 | Improved |
| add two more MLP decoder layers to LoCo (ablation) | LoCo · ImageNet-1K linear eval, 100 ep | top-1 accuracy: 66.2 -> 67.5 | Improved |
| keep only upper-module gradients on shared stage (ablation) | LoCo · ImageNet-1K linear eval, 100 ep | top-1 accuracy: 66.2 -> 65.3 | Hurt |
| replace hard sharing with L2 soft sharing (ablation) | LoCo · ImageNet-1K linear eval, 100 ep | top-1 accuracy: 66.2 -> 66.0 | Hurt |
| share only 1 block per stage instead of 1 stage (ablation) | LoCo · ImageNet-1K linear eval, 100 ep | top-1 accuracy: 66.2 -> 64.8 | Hurt |
| use Progressive ResNet-50 instead of ResNet-50 for local learning with LoCo | LoCo · ImageNet-1K linear eval, 100 ep | memory saving ratio: 1.28x -> 2.76x | Improved |
Closest papers by meaning
- Supervised Contrastive Learning (NeurIPS 2020 · similarity 0.86)
- Propagate Yourself: Exploring Pixel-Level Consistency for Unsupervised Visual Representation Learning (CVPR 2021 · similarity 0.86)
- The Loss Is Not Enough: Sampling Conditions and Inductive Bias in Contrastive Representation Learning (ICML 2026 · similarity 0.86)
- Multi-Scale and Cross-Scale Contrastive Learning for Semantic Segmentation (ECCV 2022 · similarity 0.86)
- Towards Interpretable Deep Local Learning with Successive Gradient Reconciliation (ICML 2024 · similarity 0.85)
- BECLR: Batch Enhanced Contrastive Few-Shot Learning (ICLR 2024 · 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.
