Data and Code for: Comparative Advantage of Humans vs AI in the Long Tail

Abstract: Machine learning algorithms now exceed human performance on a number of predictive tasks, generating concerns about widespread job displacement. However, supervised learning approaches rely on large amounts of high-quality labeled data and are designed for specific predictive tasks. Thus, humans may be required for a large number of tasks each of which are not commonly encountered -- the long tail -- because humans can make predictions for a broader range of outcomes and with exposure to much less data. We show that a self-supervised algorithm for chest X-rays, which does not require specifically annotated disease labels, closes this gap even in the long tail of diseases.

摘要：目前，机器学习算法（machine learning algorithms）在多项预测任务上的表现已超越人类，由此引发了关于大规模岗位替代的广泛担忧。然而，监督学习（supervised learning）方法依赖海量高质量标注数据，且专为特定预测任务设计。因此，对于大量不常出现的任务——即所谓的「长尾」任务——人类仍不可或缺，因为人类能够在接触极少数据的前提下，对更广泛的结果范围做出预测。我们的研究表明，针对胸部X光影像（chest X-rays）的自监督算法（self-supervised algorithm）无需专门标注疾病标签，即便在疾病长尾场景下，也能填补这一差距。