Bilevel Optimization under Unbounded Smoothness: A New Algorithm and Convergence Analysis

Authors: Jie Hao, Xiaochuan Gong, Mingrui Liu

ICLR 2024 | Conference PDF | Archive PDF | Plain Text | LLM Run Details

Reproducibility Variable Result LLM Response
Research Type Experimental Our experiments on hyperrepresentation learning, hyperparameter optimization, and data hyper-cleaning for text classification tasks demonstrate the effectiveness of our proposed algorithm.
Researcher Affiliation Academia Jie Hao, Xiaochuan Gong, Mingrui Liu Department of Computer Science, George Mason University, Fairfax, VA 22030, USA {jhao6, xgong2, mingruil}@gmu.edu
Pseudocode Yes Algorithm 1: BO-REP, Algorithm 2: Update Lower, Algorithm 3: Epoch-SGD
Open Source Code Yes The code is available at https://github.com/Mingrui Liu-ML-Lab/ Bilevel-Optimization-under-Unbounded-Smoothness.
Open Datasets Yes Hyper-representation experiment is conducted over Amazon Review Dataset (Blitzer et al., 2006), The hyper-cleaning experiments are conducted over the Sentiment140 dataset (Go et al., 2009) for text classification
Dataset Splits Yes Dval and Dtr denote validation and training sets respectively.
Hardware Specification Yes To accurately evaluate each algorithm, we use the machine learning framework Py Torch 1.13 to run each algorithm individually on an NVIDIA RTX A6000 graphics card, and record its training and test loss.
Software Dependencies Yes we use the machine learning framework Py Torch 1.13
Experiment Setup Yes We use grid search to tune the lower-level and upper-level step sizes from {0.001, 0.005, 0.01, 0.05, 0.1, 0.5} for all methods. The best combinations of lower-level and upper-level learning rates are (0.05, 0.1) for MAML and ANIL, (0.05, 0.05) for Stoc Bio, (0.1, 0.01) for TTSA, (0.05, 0.05) for SOBA and SABA, (0.05, 0.1) for MA-SOBA, and (0.001, 0.01) for BO-REP.