Improved Regret Bounds for Non-Convex Online-Within-Online Meta Learning

Authors: Jiechao Guan, Hui Xiong

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

Reproducibility Variable Result LLM Response
Research Type Experimental We conduct k-center clustering experiment to verify the convergence performance of task-averaged regret RT,m/m of our Algorithm 1 for non-convex OWO meta learning.
Researcher Affiliation Academia Jiechao Guan1, Hui Xiong1, 2, 1AI Thrust, The Hong Kong University of Science and Technology (Guangzhou), China 2Guangzhou HKUST Fok Ying Tung Research Institute, China {jiechaoguan, xionghui}@hkust-gz.edu.cn
Pseudocode Yes Algorithm 1 Non-convex OWO meta learning algorithm for bounded piecewise Lipschitz functions.
Open Source Code No The paper does not provide any statement or link regarding the availability of open-source code for the described methodology.
Open Datasets Yes On the one hand, we create a Gaussian mixture binary classification dataset, where each class is a 2-dimensional diagonal Guassian distribution with variance σ and 2σ, as well as the expectation (0, 0) and (bσ, 0). On the other hand, we utilize the split of the real-world Omniglot dataset to create clustering tasks, by drawing random samples each composed of five characters among which four are constant throughout. We set the number T [1, 10] of training tasks and the number m [5, 50] of samples per task for online optimization. Analogous to Balcan et al. (2021), we set the parameters γ = η = 0.01 (not hyper-parameter searched), and set the step size λ in EWA algorithm to minimize the regret in Eq. (2) (not meta-learned).
Dataset Splits No The paper specifies the number of training tasks and samples per task but does not explicitly provide train/validation/test dataset splits.
Hardware Specification No The paper does not provide any specific details about the hardware used for running its experiments.
Software Dependencies No The paper mentions software like Python in the appendix but does not provide specific version numbers for any key software components or libraries.
Experiment Setup Yes Analogous to Balcan et al. (2021), we set the parameters γ = η = 0.01 (not hyper-parameter searched), and set the step size λ in EWA algorithm to minimize the regret in Eq. (2) (not meta-learned).