Learning-to-learn non-convex piecewise-Lipschitz functions

Authors: Maria-Florina F. Balcan, Mikhail Khodak, Dravyansh Sharma, Ameet Talwalkar

NeurIPS 2021 | Conference PDF | Archive PDF | Plain Text | LLM Run Details

Reproducibility Variable Result LLM Response
Research Type Experimental We demonstrate the utility of our bounds in a series of applications across data-driven algorithm design and robust learning. This section focuses on the former and demonstrates how our results imply guarantees for meta-learning the tuning of solvers for difficult combinatorial problems. We also demonstrate the practical utility of our approach for tuning clustering on real and synthetic datasets. Next we demonstrate our meta-initialization algorithm empirically on knapsack and k-center clustering. We design experiments on real and simulated data that show the usefulness of our techniques in learning a sequence of piecewise-Lipschitz functions.
Researcher Affiliation Collaboration Maria-Florina Balcan, Mikhail Khodak, Dravyansh Sharma Carnegie Mellon University {ninamf,mkhodak,dravyans}@cs.cmu.edu Ameet Talwalkar Carnegie Mellon University & Hewlett Packard Enterprise talwalkar@cmu.edu
Pseudocode Yes Algorithm 1 Exponential Forecaster, Algorithm 2 Follow-the-Regularized-Leader (prescient form), Algorithm 3 Meta-learning the parameters of the exponential forecaster (Algorithm 1).
Open Source Code Yes Did you include the code, data, and instructions needed to reproduce the main experimental results (either in the supplemental material or as a URL)? [Yes] Supplemental material.
Open Datasets Yes We use the small split of the Omniglot dataset [33], and create clustering tasks by drawing random samples consisting of five characters each, where four characters are constant throughout.
Dataset Splits No The paper mentions training and test tasks but does not specify validation splits or detailed percentages for data partitioning.
Hardware Specification Yes Experiments run on personal computer (16GB, 2.3 GHz Dual-Core).
Software Dependencies No The paper does not provide specific software names with version numbers, or self-contained solver versions, necessary for replication.
Experiment Setup Yes We perform meta-initialization with parameters γ = η = 0.01 (no hyperparameter search performed). The step-size is set to minimize the regret term in Theorem 2.1, and not meta-learned.