Time-Independent Information-Theoretic Generalization Bounds for SGLD

Authors: Futoshi Futami, Masahiro Fujisawa

NeurIPS 2023 | Conference PDF | Archive PDF | Plain Text | LLM Run Details

Reproducibility Variable Result LLM Response
Research Type Theoretical We provide novel information-theoretic generalization bounds for stochastic gradient Langevin dynamics (SGLD) under the assumptions of smoothness and dissipativity, which are widely used in sampling and non-convex optimization studies. Our bounds are time-independent and decay to zero as the sample size increases, regardless of the number of iterations and whether the step size is fixed. Unlike previous studies, we derive the generalization error bounds by focusing on the time evolution of the Kullback Leibler divergence, which is related to the stability of datasets and is the upper bound of the mutual information between output parameters and an input dataset. Additionally, we establish the first information-theoretic generalization bound when the training and test loss are the same by showing that a loss function of SGLD is sub-exponential. This bound is also time-independent and removes the problematic step size dependence in existing work, leading to an improved excess risk bound by combining our analysis with the existing non-convex optimization error bounds.
Researcher Affiliation Academia Futoshi Futami Osaka University / RIKEN AIP futami.futoshi.es@osaka-u.ac.jp Masahiro Fujisawa RIKEN AIP masahiro.fujisawa@riken.jp
Pseudocode No The paper is theoretical and focuses on mathematical derivations and proofs; it does not contain any structured pseudocode or algorithm blocks.
Open Source Code No The paper does not provide any concrete access to source code, such as a repository link or an explicit statement about code release.
Open Datasets No The paper is theoretical and does not use or provide concrete access information for a specific publicly available or open dataset for training purposes.
Dataset Splits No The paper is theoretical and does not conduct experiments, therefore it does not provide specific dataset split information for validation.
Hardware Specification No The paper is theoretical and does not describe any specific hardware used for running experiments.
Software Dependencies No The paper is theoretical and does not mention any specific ancillary software details with version numbers required for replication.
Experiment Setup No The paper is theoretical and does not describe any specific experimental setup details, such as hyperparameter values or training configurations.