利用無監督式環境多樣化解決分佈外泛化與演算法公平性之問題

Abstract

神經網路經常透過「經驗風險最小化」進行訓練；然而，近來許多研究表明模型可能無意間學習到資料中的偏差，影響分佈外泛化能力和演算法公平性。為了解決這個問題，「不變學習」被提出用於提取不受分佈變化影響的不變特徵，而本研究提出了一個不變學習框架 EDNIL，其包含多頭神經網路以吸收資料中的偏差。在論文中，我們說明此框架不需要事先取得環境標籤，也沒有過度要求預訓練模型。此外，我們揭示了此演算法與近期討論變異特徵和不變特徵屬性研究之間的理論連結。最後，透過實驗證明，使用EDNIL訓練的模型不僅表現出卓越的分佈外泛化能力，還能促進機器學習演算法的公平性。

Neural networks are often trained with Empirical Risk Minimization (ERM). However, it has been shown that model could inadvertently learn biases present in data, causing out-of-distribution generalization and algorithmic fairness concerns. On this issue, a research direction, invariant learning, has been proposed to extract invariant features insensitive to the distributional changes. This work proposes EDNIL, an invariant learning framework containing an auxiliary multi-head neural network to absorb data biases. We show that this framework does not require prior knowledge about environments or strong assumptions about the pre-trained model. We also reveal that the proposed algorithm has theoretical connections to recent studies discussing properties of variant and invariant features. Finally, we empirically demonstrate that models trained with EDNIL not only exhibit enhanced out-of-distribution generalization but also promote algorithmic fairness.