Affine Invariant Analysis of Frank-Wolfe on Strongly Convex Sets

Authors: Thomas Kerdreux, Lewis Liu, Simon Lacoste-Julien, Damien Scieur

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

Reproducibility Variable Result LLM Response
Research Type Experimental In this work, we introduce new structural assumptions on the problem (such as the directional smoothness) and derive an affine invariant, normindependent analysis of Frank-Wolfe. We show that our rates are better than any other known convergence rates of FW in this setting. ... We conclude in Section 8 with numerical experiments.
Researcher Affiliation Collaboration 1Zuse Institute, Berlin 2D epartement d informatique et de recherche op erationnelle (DIRO), Universit e de Montr eal 3Mila, Montr eal 4Samsung SAIT AI Lab, Montr eal 5Canada CIFAR AI Chair.
Pseudocode Yes Algorithm 1 Frank-Wolfe Algorithm Input: x0 C. ... Algorithm 2 Affine invariant backtracking Input: FW vertex vk, point xk, directional smoothness estimate Lk, function f.
Open Source Code No The paper does not contain any explicit statements about releasing source code for the described methodology, nor does it provide a link to a code repository.
Open Datasets Yes Specifically, we compare our affine invariant backtracking method in Algorithm 2 against the naive FW Algorithm 1 with step size 1/L (Demyanov & Rubinov, 1970) and back-tracking FW (Pedregosa et al., 2020) on the Madelon dataset (Guyon et al., 2007).
Dataset Splits No No specific percentages or counts for training, validation, or test splits were provided. The paper only mentions using the Madelon dataset.
Hardware Specification No No specific hardware details (e.g., GPU/CPU models, memory amounts, or detailed computer specifications) used for running experiments were provided.
Software Dependencies No No specific ancillary software details (e.g., library or solver names with version numbers) were provided.
Experiment Setup Yes Here we adopt the ℓ2-ball, defined as C = {x : x 2 R}, R > 0. Specifically, we compare our affine invariant backtracking method in Algorithm 2 against the naive FW Algorithm 1 with step size 1/L (Demyanov & Rubinov, 1970) and back-tracking FW (Pedregosa et al., 2020) on the Madelon dataset (Guyon et al., 2007). The results are shown in Figure 2. In detail, we set R such that the unconstrained optimum x satisfies x 2 = 1.1R, and the initial iterate x0 = 0.