Research
152d agoBalancing Synthetic and Real Data for Generalization
Researchers at Apple unveil a theoretical framework that quantifies how much synthetic data can replace real data without hurting performance. By linking algorithmic stability to Wasserstein distance, the study identifies an optimal synthetic‑to‑real ratio that minimizes test error.
The Signal
This insight guides global AI teams building robust models when real data is limited, enhancing generalization worldwide.