NeurIPS 2019 Reproducibility Challenge Report on Ginart et al's "Making AI Forget You: Data Deletion in Machine Learning"

Jenisha Patel, Edward Son, Li Zhang

Abstract: In the context of the NeurIPS 2019 Reproducibility Challenge's Baselines Track, the baseline algorithm of Ginart et al.'s "Making AI Forget You: Data Deletion in Machine Learning" was re-implemented. Ginart et al. proposed two algorithms to support efficient data deletion for k-means clustering. Because the baseline was Lloyd's algorithm initialized by k-means++ seeding (k-means), we proceeded to re-implement the baseline. We also implemented three additional baselines, including bisecting k-means , weighted k-means and Gaussian mixture model. Among our implemented baselines, our Lloyd's algorithm initialized by k-means++ performed the best. It achieved similar clustering quality as the baseline implemented by Ginart et al. However, our implementation of the baseline lead to lower amortized runtimes, which can be attributed to a more computationally optimal implementation.

Track: Baseline

NeurIPS Paper Id: https://openreview.net/forum?id=H1xovVBg8B