Research map / Label-efficient and robust learning / Recognition and long-tail
Matched Data, Better Models: Target Aligned Data Filtering with Sparse Features
Arnav Mohanty Das, Gantavya Bhatt, Sahil Verma, Yiping Wang, Viswa Virinchi Muppirala, Jeff Bilmes
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 sparse SAE features instead of dense CLIP features (ablation) | SDM · DataComp-medium ImageNet-1K | accuracy: 24.1 -> 35.2 | Improved |
| use sparse SAE features instead of dense CLIP features (ablation) | SDM · DataComp-medium 38-task average | accuracy: 29.1 -> 36.4 | Improved |
| add concave log term (submodularity) (ablation) | SDM · DataComp-medium 38-task average | accuracy: 35.1 -> 36.4 | Improved |
| add concave log term (submodularity) (ablation) | SDM · DataComp-medium ImageNet-1K | accuracy: 34.6 -> 35.2 | Improved |
| add monotonicity loss L_mono to SAE training (ablation) | SDM · DataComp-medium 38-task average | accuracy: 35.00% -> 36.40% | Improved |
| add monotonicity loss L_mono to SAE training (ablation) | SDM · DataComp-medium ImageNet-1K | accuracy: 34.80% -> 35.20% | Improved |
| intersect 5 stochastic greedy runs instead of a single run (ablation) | SDM · DataComp-medium ImageNet-1K | accuracy: 33.5 -> 35.2 | Improved |
| use a broader 24-dataset target instead of ImageNet-1K (ablation) | SDM · DataComp-medium 38-task average | accuracy: 36.4 -> 35.8 | Hurt |
| use SDM instead of NCL ∩ NormSim | NCL ∩ NormSim (IN1K) · DataComp-medium ImageNet-1K | accuracy: 32.8% -> 35.2% | Improved |
| use SDM instead of NCL ∩ NormSim | NCL ∩ NormSim (Target) · DataComp-medium 38-task average | accuracy: 35.7% -> 36.4% | Improved |
| combine SDM with NCL + DFN + HYPE | NCL + DFN + HYPE · DataComp-medium 38-task average | accuracy: 38.8% -> 39.2% | Improved |
| combine SDM with NCL + DFN + HYPE | NCL + DFN + HYPE · DataComp-medium ImageNet-1K | accuracy: 38.2% -> 39.2% | Improved |
Closest papers by meaning
- Data Filtering Networks (ICLR 2024 · similarity 0.86)
- SELECT: A Large-Scale Benchmark of Data Curation Strategies for Image Classification (NeurIPS 2024 · similarity 0.85)
- Diffusion Model for Dense Matching (ICLR 2024 · similarity 0.85)
- Concept-Aware Batch Sampling Improves Language-Image Pretraining (CVPR 2026 · similarity 0.85)
- Filter Like You Test: Data-Driven Data Filtering for CLIP Pretraining (NeurIPS 2025 · similarity 0.85)
- A Unified Debiasing Approach for Vision-Language Models across Modalities and Tasks (NeurIPS 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.
