基於深度學習之群眾外包資料學習問題

Abstract

由於深度學習方法對於標記資料的需求較大，透過群眾外包獲取訓練數據成為了一種方便且具有經濟效益的方式。為了處理收集到的標籤中的雜訊，近年來出現了許多彙整標籤並學習分類器的演算法。這些演算法往往會受到專家偏見(expert bias)以及數據偏差(data bias)的影響。具體而言，每個專家可能具有不同的背景和能力，導致他們犯下特定類型的錯誤，另外資料本身的特徵也可能系統性地對全體專家產生影響。我們基於深度學習，提出了一個可以同時推斷專家偏見以及數據偏差並學習分類器的模型。此外，我們也透過實驗證明提出的方法在模擬的資料以及真實世界資料的不同情境中都有良好的效果。

Due to the large amount of labeled data required for deep learning methods, crowdsourcing has emerged as a convenient and cost-efficient way of obtaining training data. To deal with the noisy nature of the collected labels, many algorithms have been proposed to infer true labels and learn a classifier from the obtained noisy labels. The algorithms are often challenged by the occurrence of expert bias and data bias. Specifically, each expert may have diverse backgrounds and abilities, causing them to make specific types of mistakes, and each example from a given data may have features that systematically shift the majority opinion of experts. We propose a model based on deep learning that simultaneously infers the expert bias, the data bias, and learns a classifier. We empirically show the effectiveness of our model on both simulated and real world datasets under different scenarios.