This repo contains the code for the numerical results presented in the following papers.
1- "Information-Theoretic Generalization Bounds for SGLD via Data-Dependent Estimates"
Published at NeurIPS'19
[https://arxiv.org/abs/1911.02151]
by Jeffrey Negrea*, Mahdi Haghifam*, Gintare Karolina Dziugaite, Ashish Khisti, Daniel M. Roy
2- "Sharpened Generalization Bounds based on Conditional Mutual Information and an Application to Noisy, Iterative Algorithms"
Published at NeurIPS'20
[https://arxiv.org/abs/2004.12983]
by Mahdi Haghifam, Jeffrey Negrea, Ashish Khisti, Daniel M Roy, Gintare Karolina Dziugaite
We only consider the full-batch SGLD, i.e., LD, for simplicity here.