基於部分標記環境下的不變式學習

Abstract

實現風險穩健的分佈外泛化(OOD)仍然是透過經驗最小化(ERM)訓練的機器學習模型面臨的巨大挑戰。儘管不變式最小化(IRM)提供了一種有前景的方法，但它要求對訓練資料進行環境劃分，造成實際使用上的困難。最近出現使用無環境資料的方法雖然有效，但往往依賴假設或輔助資訊。我們提出了EEPL，這是一種新穎的方法，透過部分預定義環境來產生高品質的增強環境。EEPL克服了理論上的限制，並在複雜的場景中實現了超越現有方法的最佳表現。此外，我們將「多樣性」確定為有效不變學習的關鍵屬性。我們的研究為實現可靠的OOD泛化開拓了新的研究途徑。

Achieving robust out-of-distribution (OOD) generalization remains a challenge for machine learning models trained with Empirical Risk Minimization (ERM). While Invariant Risk Minimization (IRM) offers a promising approach, it requires impractical environment partitioning of the training data. Recent environment-free methods have emerged, but they often rely on assumptions or auxiliary information. We propose EEPL, a novel approach that leverages a small set of pre-defined environments to generate high-quality augmented environments. EEPL overcomes theoretical limitations and achieves state-of-the-art performance, surpassing existing methods in challenging scenarios. Additionally, we verify ``diversity'' as a crucial property for effective invariant learning. Our work opens new avenues for research in achieving reliable OOD generalization.