We investigate replicable learning algorithms. Ideally, we would like to design algorithms that output the same canonical model over multiple runs, even when different runs observe a different set of samples from the unknown data distribution. In general, such a strong notion of replicability is not achievable. Thus we consider two feasible notions of replicability called list replicability and certificate replicability. Intuitively, these notions capture the degree of (non) replicability. We design algorithms for certain learning problems that are optimal in list and certificate complexity. We establish matching impossibility results.
翻译:我们研究可复制的学习算法。理想情况下,我们希望设计算法,即使在不同的运行中观察到未知数据分布的不同样本集,也能输出相同的规范模型。通常情况下,这种强复制性的概念是无法实现的。因此,我们考虑两种可行的复制性概念,称为列表复制性和证书复制性。感性地说,这些概念捕捉了(非)可复制性的程度。我们为某些学习问题设计了最佳的列表和证书复杂度算法,并建立了匹配不可能结果。