Joint Dimensionality Reduction and Metric Learning: A Geometric Take
Authors: Mehrtash Harandi, Mathieu Salzmann, Richard Hartley
ICML 2017 | Conference PDF | Archive PDF | Plain Text | LLM Run Details
| Reproducibility Variable | Result | LLM Response |
|---|---|---|
| Research Type | Experimental | Our experiments evidence that, while we directly work on high-dimensional features, our approach yields competitive runtimes with and higher accuracy than state-of-the-art metric learning algorithms. ... 6. Experimental Evaluation We now evaluate our algorithms (DRML and k DRML) and compare them with the representative baseline metric learning methods discussed above, i.e., NCA (Goldberger et al., 2004) LMNN (Weinberger & Saul, 2009), ITML (Davis et al., 2007), LDML (Guillaumin et al., 2009), KISSME (Koestinger et al., 2012) and GMML (Zadeh et al., 2016), as well as with dataset-speciļ¬c baselines mentioned below. |
| Researcher Affiliation | Academia | 1Data61, CSIRO, Canberra, Australia 2Australian National University, Canberra, Australia 3CVLab, EPFL, Switzerland. |
| Pseudocode | No | The paper describes the optimization steps but does not include a structured pseudocode block or an algorithm formally labeled as such. |
| Open Source Code | Yes | In our experiments, we employed Conjugate Gradient descent on M to solve (12). In particular, we implemented the operations required for our manifold within the manopt Riemannian optimization toolbox (Boumal et al., 2014). The code is available at https://sites.google.com/site/mehrtashharandi/. |
| Open Datasets | Yes | ASLAN dataset (Kliper Gross et al., 2012) ... i LIDS dataset (Zheng et al., 2009) ... Youtube Faces (YTF) dataset (Wolf et al., 2011) |
| Dataset Splits | No | For the ASLAN dataset, the paper mentions '5,400 training and 600 testing pairs'. For the i LIDS dataset, it states 'randomly divided into two subsets, training and test'. For the YTF dataset, it mentions '10 folds' for video pairs. None of these explicitly define a separate 'validation' split. |
| Hardware Specification | No | The paper does not provide specific details on the hardware used for running the experiments. |
| Software Dependencies | No | The paper mentions using 'the manopt Riemannian optimization toolbox (Boumal et al., 2014)' but does not specify a version number for it or any other software dependencies. |
| Experiment Setup | Yes | In all our experiments, we followed the so-called restricted protocol. That is, the only information accessible to the algorithms is the similarity/dissimilarity labels of pairs of samples; the class labels of the samples are unknown. For all the methods, we report the results obtained with the best subspace dimension. ... For k DRML, we used an RBF Gaussian kernel whose bandwidth was set using Jaakkola s heuristic (Jaakkola et al., 1999). |