关于 #今日arXiv精选 

这是「AI 学术前沿」旗下的一档栏目，编辑将每日从arXiv中精选高质量论文，推送给读者。

Image In painting Applied to Art Completing Escher's Print Gallery

Comment: Abstract submitted to LatinX workshop on ICML 2021

Link: http://arxiv.org/abs/2109.02536

Abstract

This extended abstract presents the first stages of a research on in-paintingsuited for art reconstruction. We introduce M.C Eschers Print Gallerylithography as a use case example. This artwork presents a void on its centerand additionally, it follows a challenging mathematical structure that needs tobe preserved by the in-painting method. We present our work so far and ourfuture line of research.

Statistical Privacy Guarantees of Machine Learning Preprocessing Techniques

Comment: Accepted to the ICML 2021 Theory and Practice of Differential Privacy  Workshop

Link: http://arxiv.org/abs/2109.02496

Abstract

Differential privacy provides strong privacy guarantees for machine learningapplications. Much recent work has been focused on developing differentiallyprivate models, however there has been a gap in other stages of the machinelearning pipeline, in particular during the preprocessing phase. Ourcontributions are twofold: we adapt a privacy violation detection frameworkbased on statistical methods to empirically measure privacy levels of machinelearning pipelines, and apply the newly created framework to show thatresampling techniques used when dealing with imbalanced datasets cause theresultant model to leak more privacy. These results highlight the need fordeveloping private preprocessing techniques.

On Second-order Optimization Methods for Federated Learning

Comment: ICML 2021 Workshop "Beyond first-order methods in ML systems"

Link: http://arxiv.org/abs/2109.02388

Abstract

We consider federated learning (FL), where the training data is distributedacross a large number of clients. The standard optimization method in thissetting is Federated Averaging (FedAvg), which performs multiple localfirst-order optimization steps between communication rounds. In this work, weevaluate the performance of several second-order distributed methods with localsteps in the FL setting which promise to have favorable convergence properties.  We (i) show that FedAvg performs surprisingly well against its second-ordercompetitors when evaluated under fair metrics (equal amount of localcomputations)-in contrast to the results of previous work. Based on ournumerical study, we propose (ii) a novel variant that uses second-order localinformation for updates and a global line search to counteract the resultinglocal specificity.

·