A Fully First-Order Method for Stochastic Bilevel Optimization

Authors: Jeongyeol Kwon, Dohyun Kwon, Stephen Wright, Robert D Nowak

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

Reproducibility Variable Result LLM Response
Research Type Experimental We demonstrate even superior practical performance of the proposed method over existing second-order based approaches on MNIST data-hypercleaning experiments. We demonstrate the proposed algorithms on a data hyper-cleaning task involving MNIST (Deng, 2012).
Researcher Affiliation Academia 1University of Wisconsin-Madison, USA 2University of Seoul, Korea.
Pseudocode Yes Algorithm 1 F2SA and Algorithm 2 F3SA are included in the paper.
Open Source Code No The paper does not contain any explicit statements or links indicating the availability of open-source code for the described methodology.
Open Datasets Yes We demonstrate the proposed algorithms on a data hyper-cleaning task involving MNIST (Deng, 2012).
Dataset Splits Yes We are given a noisy training set Dtrain := {( xi, yi)}n i=1 with the label yi being randomly corrupted with probability p < 1. We are also given a small but clean validation set Dval := {(xi, yi)}m i=1. ... We use n = 19000 training samples and m = 1000 clean validation samples...
Hardware Specification No The paper does not provide any specific hardware details such as GPU/CPU models or memory specifications used for experiments.
Software Dependencies No The paper does not list specific software dependencies with version numbers (e.g., Python 3.x, PyTorch 1.x).
Experiment Setup Yes We use n = 19000 training samples and m = 1000 clean validation samples with regularization parameter c = 0.01. We demonstrate the performance of Algorithm 1 (F2SA) and the second-order based method (SOBO) with batch sizes 50 and 500.