The Day Synthetic Data Turned Poisonous: Inside Model Collapse
中文摘要
递归使用合成数据训练会导致“模型崩溃”，这会消除多样性、放大错误，并使生成式模型逐渐脱离现实。
English Summary
Recursive training on synthetic data causes "model collapse," leading to diminished diversity, amplified errors, and a disconnect from reality as models learn from their own outputs.
原文节选
Why recursive training loops silently erase diversity, amplify errors, and push generative models away from reality and why even one real… Continue reading on Towards AI »