SOSELETO: A Unified Approach to Transfer Learning and Training with Noisy Labels


Author
Or Litany, Daniel Freedman
Published Year
2018
Publisher
Carnegie Mellon University
Abstract
We present SOSELETO (SOurce SELEction for Target Optimization), a new methodfor exploiting a source dataset to solve a classification problem on a targetdataset. SOSELETO is based on the following simple intuition: some sourceexamples are more informative than others for the target problem. To capturethis intuition, source samples are each given weights; these weights are solvedfor jointly with the source and target classification problems via a bileveloptimization scheme. The target therefore gets to choose the source sampleswhich are most informative for its own classification task. Furthermore, thebilevel nature of the optimization acts as a kind of regularization on thetarget, mitigating overfitting. SOSELETO may be applied to both classictransfer learning, as well as the problem of training on datasets with noisylabels; we show state of the art results on both of these problems.