Generalised Mutual Information for Discriminative Clustering
Authors: Louis Ohl, Pierre-Alexandre Mattei, Charles Bouveyron, Warith HARCHAOUI, Mickaël Leclercq, Arnaud Droit, Frederic Precioso
NeurIPS 2022 | Conference PDF | Archive PDF | Plain Text | LLM Run Details
| Reproducibility Variable | Result | LLM Response |
|---|---|---|
| Research Type | Experimental | For all experiments below, we report the adjusted rand index (ARI) (Hubert and Arabie, 1985), a common metric in clustering. This metric is external as it requires labels for evaluation. It ranges from 0, when labels are independent from cluster assignments, to 1, when labels are equivalent to cluster assignments up to permutations. An ARI close to 0 is equivalent to the best accuracy when voting constantly for the majority class, e.g. 10% on a balanced 10-class dataset. Regarding the MMDand Wasserstein-GEMINIs, we used by default a linear kernel and the Euclidean distance unless specified otherwise. All discriminative models are trained using the Adam optimiser (Kingma and Ba, 2014). We estimate a total of 450 hours of GPU consumption. (See Appendix I for the details of Python packages for experiments and Appendix F for further experiments regarding model selection). The code is available at https://github.com/oshillou/GEMINI |
| Researcher Affiliation | Collaboration | Louis Ohl Université Côte d Azur Inria, CNRS I3S, Maasai team CHU de Québec Research Center Laval University louis.ohl@inria.fr Pierre-Alexandre Mattei Université Côte d Azur Inria, CNRS LJAD, Maasai team pierre-alexandre.mattei@inria.fr Charles Bouveryon Université Côte d Azur Inria, CNRS LJAD, Maasai team Warith Harchaoui Jellysmack AI Labs Research and Development Mickael Leclerq CHU de Québec Research Center Laval University Arnaud Droit CHU de Québec Research Center Laval University Frederic Precioso Université Côte d Azur Inria, CNRS I3S, Maasai team |
| Pseudocode | No | The paper does not contain a pseudocode block or algorithm clearly labeled as such. |
| Open Source Code | Yes | The code is available at https://github.com/oshillou/GEMINI |
| Open Datasets | Yes | We trained a neural network using either MI or GEMINIs. Following Hu et al. (2017), we first tried with a MLP with one single hidden layer of dimension 1200. To further illustrate the robustness of the method and its adaptability to other architectures, we also experimented using a Le Net-5 architecture (Le Cun et al., 1998) since it is adequate to the MNIST dataset. We report our results in Table 3. |
| Dataset Splits | No | The paper uses standard datasets like MNIST and CIFAR10 but does not specify the percentages or counts for training, validation, and test splits needed for reproduction. It only mentions 'Data generation processes are reported in supplementary materials and main parameters are described at the head of Sec. 4' in the checklist. |
| Hardware Specification | No | The paper states: 'We estimate a total of 450 hours of GPU consumption.' However, it does not specify the exact GPU models, CPU types, or other hardware components used for the experiments. |
| Software Dependencies | Yes | All packages used in this paper for the experiments are listed below with their version: Pytorch (1.10.0), Scikit-Learn (1.0.2), Pandas (1.3.5), Num Py (1.21.5), Matplotlib (3.5.1), POT (0.8.2) |
| Experiment Setup | Yes | For all experiments below, we report the adjusted rand index (ARI) (Hubert and Arabie, 1985), a common metric in clustering. ... All discriminative models are trained using the Adam optimiser (Kingma and Ba, 2014). ... We trained a neural network using either MI or GEMINIs. ... We also experimented using a Le Net-5 architecture (Le Cun et al., 1998) ... The ARI score of models trained for 200 epochs on CIFAR10 ... for MNIST for 500 epochs. |