Research map / Learning theory and algorithms / Prediction and decision losses
Data Shapley in One Training Run
Read the paperCheck your idea against 108,967 papers
AI-extracted summary. Written by AI from the paper's full text. Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.
Proposes In-Run Data Shapley, which computes exact-in-approximation Shapley values for each SGD update within a single training run via first/second-order Taylor expansion, with "ghost" dot-product techniques that make it nearly as fast as regular training.
Problem
Data Shapley requires retraining models on many data subsets, which is infeasible for foundation models and measures contribution to a learning algorithm rather than to a specific training run.
Approach
Decompose the utility into per-iteration local loss changes, approximate them with Taylor expansions so the Shapley value has closed form (gradient dot-products and gradient-Hessian-gradient products), and compute these efficiently in one or two backward passes using ghost-clipping-style tricks, then sum across the training run.
Key results
- Ghost first-order In-Run Data Shapley runs at 70.5 vs 76.2 points/s for regular training on GPT2-Small (single A100), while the direct implementation runs at 4.2.
- Removing negatively valued Pile data yields about 25% fewer iterations to reach test loss 3.75; about 16% of corpora had negative second-order values.
- First-order approximation vs 1000-permutation Monte Carlo ground truth gives RMSE around 0.0003 at iteration 3500 on GPT2.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Use first-order In-Run Data Shapley with ghost dot-product instead of regular training | GPT2-Small · Single 80GB A100 | throughput: 76.2 -> 70.5 | Hurt |
| Use second-order In-Run Data Shapley with ghost computation instead of regular training | GPT2-Small · Single 80GB A100 | throughput: 76.2 -> 34.4 | Hurt |
| Use ghost dot-product rather than the direct implementation | First-order In-Run Data Shapley · GPT2-Small, single 80GB A100 | throughput: 4.2 -> 70.5 | Helped |
| Use ghost computation rather than the direct implementation | Second-order In-Run Data Shapley · GPT2-Small, single 80GB A100 | throughput: 1.8 -> 34.4 | Helped |
| Use a first-order Taylor approximation (ablation) | In-Run Data Shapley · GPT2, Pile, 3500th training iteration | RMSE against Monte Carlo estimate: 0.0003 | Helped |
| Remove negatively valued data points using In-Run Data Shapley | GPT2 trained on a Pile subset · Pile subset, test loss 3.75 | training iterations to reach test loss: around 25% fewer | Helped |
| Use first-order rather than second-order approximation (ablation) | In-Run Data Shapley · GPT2, Pile, 3500th training iteration | RMSE against Monte Carlo estimate: 0.0003 | No clear effect |
| Use In-Run Data Shapley rather than influence function to filter low-quality data | GPT2 trained on a Pile subset · Pile subset, test loss | test loss convergence | Helped |
| Attribute a significant paraphrase with first-order In-Run Data Shapley rather than BM25 | Data attribution methods · Significant paraphrase | average value rank: 1.6 -> 32.3 | Mixed |
| Attribute a rewritten, similar-topic corpus with In-Run Data Shapley rather than BM25 | Data attribution methods · Similar topic | average value rank: 20917.3 -> 145.6 | Helped |
Surprising findings
- About 16% of the Pile training corpora receive negative second-order In-Run Data Shapley values despite curation.
- Original training data ranks highly (top ~145 of ~320k for similar-topic validation) even when the validation text is a complete rewrite, suggesting contribution without memorization.
Limitations
- Requires validation data to be available before training.
- Currently limited to SGD; extension to Adam is not solved.
- Large batch sizes need gradient accumulation, increasing runtime.
Experimental setup
Datasets: The Pile
Benchmarks: The Pile validation set
Metrics: test loss throughput (data points per second) RMSE vs Monte Carlo Shapley average value rank Spearman correlation
Baselines: Influence function BM25 TracIN-Ideal Monte Carlo Shapley
Builds on: Data Shapley TracIN ghost clipping influence functions
Closest papers by meaning
- Data Shapley: Equitable Valuation of Data for Machine Learning (ICML 2019 · similarity 0.87)
- Helpful or Harmful Data? Fine-tuning-free Shapley Attribution for Explaining Language Model Predictions (ICML 2024 · similarity 0.86)
- shapiq: Shapley Interactions for Machine Learning (NeurIPS 2024 · similarity 0.85)
- DU-Shapley: A Shapley Value Proxy for Efficient Dataset Valuation (NeurIPS 2024 · similarity 0.85)
- Rethinking Data Shapley for Data Selection Tasks: Misleads and Merits (ICML 2024 · similarity 0.85)
- Distilled Datamodel with Reverse Gradient Matching (CVPR 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.