Research map / Federated learning and privacy / Personalized and heterogeneous FL
Collaborative Learning of Discrete Distributions under Heterogeneity and Communication Constraints
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 |
|---|---|---|---|
| Use SHIFT with trimmed-mean aggregation instead of local Minimax estimation | Shakespeare k-gram distribution estimation · Shakespeare k-grams, k=2, b=2, 4, 6, 8 | average ℓ2 error: 640 ± 6.0 -> 36 ± 2.2; 142 ± 1.2 -> 19 ± 0.51; 40 ± 0.40 -> 13 ± 0.24; 14 ± 0.13 -> 10 ± 0.062 | Helped |
| Use SHIFT with trimmed-mean aggregation instead of global Minimax pooling | Shakespeare k-gram distribution estimation · Shakespeare k-grams, k=2, b=2, 4, 6, 8 | average ℓ2 error: 33 ± 1.8 -> 36 ± 2.2; 17 ± 0.37 -> 19 ± 0.51; 14 ± 0.081 -> 13 ± 0.24; 13 ± 0.037 -> 10 ± 0.062 | Mixed |
| Use SHIFT with trimmed-mean aggregation instead of local Minimax estimation | Shakespeare k-gram distribution estimation · Shakespeare k-grams, k=3, b=2, 4, 6, 8 | average ℓ2 error: 15000 ± 21 -> 5100 ± 6.3; 3000 ± 5.9 -> 140 ± 2.3; 720 ± 2.1 -> 43 ± 0.66; 180 ± 0.39 -> 18 ± 0.18 | Helped |
| Use SHIFT with trimmed-mean aggregation instead of global Minimax pooling | Shakespeare k-gram distribution estimation · Shakespeare k-grams, k=3, b=2, 4, 6, 8 | average ℓ2 error: 4400 ± 5.7 -> 5100 ± 6.3; 100 ± 1.4 -> 140 ± 2.3; 38 ± 0.35 -> 43 ± 0.66; 23 ± 0.090 -> 18 ± 0.18 | Mixed |
| Replace entry-wise median with entry-wise trimmed mean as SHIFT's robust estimator (ablation) | SHIFT · Shakespeare k-grams, k=2, b=2, 4, 6, 8 | average ℓ2 error: 47 ± 2.4 -> 36 ± 2.2; 21 ± 0.66 -> 19 ± 0.51; 14 ± 0.17 -> 13 ± 0.24; 11 ± 0.10 -> 10 ± 0.062 | Helped |
| Replace entry-wise median with entry-wise trimmed mean as SHIFT's robust estimator (ablation) | SHIFT · Shakespeare k-grams, k=3, b=2, 4, 6, 8 | average ℓ2 error: 7300 ± 9.6 -> 5100 ± 6.3; 180 ± 2.1 -> 140 ± 2.3; 53 ± 1.0 -> 43 ± 0.66; 20 ± 0.18 -> 18 ± 0.18 | Helped |
| Pool heterogeneous data globally without per-cluster fine-tuning | Minimax global baseline · Synthetic data, varied sample size and cluster count | average ℓ2 error | Hurt |
| Increase the number of clusters used for collaboration | SHIFT · Synthetic data | average ℓ2 error | Helped |
| Increase heterogeneity until it is close to the dimension | SHIFT collaboration · Synthetic data | average ℓ2 error | Hurt |
Closest papers by meaning
- Communication-Efficient Distributed Learning of Discrete Distributions (NeurIPS 2017 · similarity 0.88)
- Robust Distributed Learning: Tight Error Bounds and Breakdown Point under Data Heterogeneity (NeurIPS 2023 · similarity 0.87)
- Collaborative Learning by Detecting Collaboration Partners (NeurIPS 2022 · similarity 0.86)
- Derandomizing Multi-Distribution Learning (NeurIPS 2024 · similarity 0.86)
- Positive Distribution Shift as a Framework for Understanding Tractable Learning (ICML 2026 · similarity 0.86)
- Collaborative Learning with Different Labeling Functions (ICML 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.