Research
151d agoSynthetic Data Balances Generalization
Apple’s new framework quantifies how much synthetic data can replace real data without hurting performance. By linking algorithmic stability to Wasserstein distance, researchers predict the optimal synthetic‑to‑real ratio that minimizes test error.
The Signal
The method offers a principled way for global teams to generate realistic datasets, improving model robustness across industries and accelerating AI deployment worldwide.