Notice: The reproducibility variables underlying each score are classified using an automated LLM-based pipeline, validated against a manually labeled dataset. LLM-based classification introduces uncertainty and potential bias; scores should be interpreted as estimates. Full accuracy metrics and methodology are described in [1].

Geometric Mean Metric Learning

Authors: Pourya Zadeh, Reshad Hosseini, Suvrit Sra

ICML 2016 | Venue PDF | LLM Run Details

Reproducibility Variable Result LLM Response
Research Type Experimental Furthermore, on standard benchmark datasets, our closed-form solution consistently attains higher classification accuracy. Validation. We consider multi-class classification using the learned metrics, and validate GMML by comparing it against widely used metric learning methods. GMML runs up to three orders of magnitude faster while consistently delivering equal or higher classification accuracy.
Researcher Affiliation Academia Pourya Habib Zadeh EMAIL Reshad Hosseini EMAIL School of ECE, College of Engineering, University of Tehran, Tehran, Iran Suvrit Sra EMAIL Massachusetts Institute of Technology, Cambridge, MA, USA
Pseudocode Yes Algorithm 1 Geometric Mean Metric Learning
Open Source Code No The paper does not provide an unambiguous statement of releasing the source code for the methodology described, nor does it provide a direct link to a source-code repository.
Open Datasets Yes The datasets are obtained from the well-known UCI repository (Asuncion & Newman, 2007). The datasets in this experiment are Isolet, Letters (Asuncion & Newman, 2007), MNIST3 (Le Cun et al., 1998) and USPS (Le Cun et al., 1990).
Dataset Splits Yes Figure 2 reports 40 runs of a two-fold splitting of the data. In each run, the data is randomly divided into two equal sets. We use five-fold cross-validation for choosing the best parameter t. Figure 4 reports the average classification error over 5 runs of random splitting of the data. We use three-fold cross-validation for adjusting the parameter t.
Hardware Specification Yes All methods were implemented on MATLAB R2014a (64-bit), and the simulations were run on a personal laptop with an Intel Core i5 (2.5Ghz) processor under the OS X Yosemite operating system.
Software Dependencies Yes All methods were implemented on MATLAB R2014a (64-bit), and the simulations were run on a personal laptop with an Intel Core i5 (2.5Ghz) processor under the OS X Yosemite operating system.
Experiment Setup Yes We choose k = 5, and estimate a full-rank matrix A in all methods. The regularization parameter λ is set to zero for most of the datasets. We only add a small value of λ when the similarity matrix S becomes singular. For example, since the similarity matrix of the Segment data is near singular, we use the regularized version of our method with λ = 0.1 and A0 equals to the identity matrix. We use five-fold cross-validation for choosing the best parameter t. In the first step the best t is chosen among the values {0.1, 0.3, 0.5, 0.7, 0.9}.