Chiosa
Machine Learning

Synthetic training can fail where generated data leave gaps

2 min
Updated
#cost-performance#ood-robustness

In plain words

This theory paper asks when a binary classifier trained on generated examples can also predict well on real data.

It analyzes training that picks a classifier with low average prediction loss on a sample of synthetic examples, assuming real and synthetic data share the same rule for assigning labels.

The theory gives a guarantee when synthetic data cover real inputs under certain bounds. If they leave gaps, a classifier can do very well on synthetic images and badly on real counterparts: one constructed example made up to 0.5% errors on synthetic images and 99.25% on real ones. That shows what can happen, not what ordinary training usually produces.

The research

When generated examples stand in for real data, can a classifier that performs well on those examples still fail on real inputs? The paper examines that question when both data sources share a rule for assigning labels but the generated examples may leave some real regions unseen.

The paper derives when synthetic training transfers, shows how uncovered real regions can defeat guarantees, and identifies conditions that restore a bound.

The paper asks when low error on generated examples tells us anything about a classifier's error on real inputs. It analyzes binary classification when real and synthetic data share a label rule but may cover different input regions.

The theory gives a finite-sample real-error bound when synthetic data cover the real distribution under a bounded density ratio. With uncovered regions, it permits classifiers that perform nearly optimally on synthetic data yet fail on real data.

Image experiments demonstrate such a constructed case, while ordinary tested architectures showed positively related errors across domains.

These guarantees depend on assumptions about the real and synthetic distributions or on Lipschitz conditions for the classifier and label rule.

The image experiments cover eight ImageNet classes, and the adversarial classifier demonstrates that a failure can exist, not that ordinary training will produce it.

Analysis

Deep

Original paperWhen Does Training on Synthetic Data Provably Work?Open original
Submitted to arXiv
Related paperMachine LearningSame-size students struggle to match model harnesses#cost-performance#policy-distillation#harness-optimizationRelated paperMachine LearningModel weight spectra rank robustness within model groups#cost-performance#weight-spectra#ood-robustnessRelated paperMachine LearningParallel language models coordinate several token predictions at once#cost-performance#parallel-decodingRelated paperMachine LearningMatching training optimizers adds little consistent benefit#cost-performance#matrix-aware-optimization
Next paperSame-size students struggle to match model harnesses