Primal and Dual Analysis of Entropic Fictitious Play for Finite-sum Problems

Authors: Atsushi Nitanda, Kazusato Oko, Denny Wu, Nobuhito Takenouchi, Taiji Suzuki

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

Reproducibility Variable Result LLM Response
Research Type Experimental We illustrate the efficiency of our novel implementation in experiments including neural network optimization and image synthesis.
Researcher Affiliation Academia 1Kyushu Institute of Technology 2Center for Advanced Intelligence Project 3University of Tokyo 4University of Toronto 5Vector Institute for Artificial intelligence.
Pseudocode Yes Algorithm 1 Discrete-time Entropic Fictitious Play; Algorithm 2 Efficient Implementation of EFP for Finite-sum Problem
Open Source Code No The paper does not provide concrete access to source code for the methodology described.
Open Datasets No The paper mentions generating data for a student-teacher setting and using the Mona Lisa image as a target, but does not provide specific access information (link, DOI, formal citation) for any publicly available datasets used for training.
Dataset Splits No The paper does not provide specific dataset split information for training, validation, or testing.
Hardware Specification No The paper does not provide specific hardware details (e.g., exact GPU/CPU models, memory amounts) used for running its experiments.
Software Dependencies No The paper does not provide specific ancillary software details with version numbers.
Experiment Setup Yes We optimize the neural network using EFP with an outerloop step size ηγ = 0.01. At each iteration, we approximate the proximal Gibbs measure ˆµt via the Langevin Monte Carlo algorithm with step size η = 0.01. (Section 7.1) We run Algorithm 2 with λ = 10 5, λ = 10 4, T = 2000, S = 10, m = 1000, η γ = 0.01 to fit the target image. As for the step size for Langevin Monte Carlo, we used cosine annealing from 0.1 to 0.01. (Figure 3 caption)