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 replic