Unsupervised Riemannian Metric Learning for Histograms Using Aitchison Transformations
Authors: Tam Le, Marco Cuturi
ICML 2015 | Conference PDF | Archive PDF | Plain Text | LLM Run Details
| Reproducibility Variable | Result | LLM Response |
|---|---|---|
| Research Type | Experimental | We provide experimental evidence that the metric obtained under our proposal outperforms alternative approaches. We provide experimental evidence in Section 7, and concluding this paper in Section 8. 7. Experiments 7.1. Clustering application with K-Medoids 7.2. k-Nearest Neighbors Classification with Locally Sensitive Hashing |
| Researcher Affiliation | Academia | Tam Le TAM.LE@IIP.IST.I.KYOTO-U.AC.JP Graduate School of Informatics, Kyoto University, Japan Marco Cuturi MCUTURI@I.KYOTO-U.AC.JP Graduate School of Informatics, Kyoto University, Japan |
| Pseudocode | Yes | Algorithm 1 Gradient Ascent using Contrastive Divergence |
| Open Source Code | No | The paper mentions using existing toolboxes (Label Me, gensim, PMTK3) but does not state that the authors are releasing their own source code for the methodology described. |
| Open Datasets | Yes | We use the K-medoids clustering algorithm... We test our method on 6 benchmark datasets. Table 1 displays their properties and parameters. These datasets include different kinds of data such as scene images in MIT Scene2 and UIUC Scene3 datasets, flower images in Oxford Flower4 dataset, object images in CALTECH-1015 dataset and texts in Reuters6 and 20 News Group7 datasets. We also carry out k-nearest neighbors classification with locally sensitive hashing. We use 2 large datasets MNIST60K12 and CIFAR-1013. |
| Dataset Splits | No | The paper describes a training/test split for the k-Nearest Neighbors experiment ("we randomly choose 50000 images as a database and use the rest 10000 images for queries") but does not explicitly mention a separate validation set for hyperparameter tuning or model selection in either the clustering or classification experiments. |
| Hardware Specification | No | The paper does not provide any specific details about the hardware (e.g., GPU models, CPU types, memory) used to run the experiments. |
| Software Dependencies | No | For image datasets... We use the Label Me toolbox8 for computing dense SIFT features. For text datasets... using the gensim toolbox9. We use the PMTK3 toolbox10 implementation of the K-medoids algorithm. The paper mentions the names of software toolboxes used but does not provide specific version numbers for these dependencies. |
| Experiment Setup | Yes | We set maximum iterations tmax = 10000 and a tolerance ϵ = 10 5. We choose gradient step size tα 0 and tλ 0 from the sets {0.001, 0.005, 0.01, 0.05, 0.1, 0.5} respectively and µ from {0.1, 1, 10}. We also set 5 cycles for Metropolis-Hasting sampling algorithm to transform training data into data drawn from p(x). |