Deep Embedded Non-Redundant Clustering

Authors: Lukas Miklautz, Dominik Mautz, Muzaffer Can Altinigneli, Christian Böhm, Claudia Plant5174-5181

AAAI 2020 | Conference PDF | Archive PDF | Plain Text | LLM Run Details

Reproducibility Variable Result LLM Response
Research Type Experimental We evaluated ENRC with four different data sets. As we are the first to address the topic of non-redundant clustering with neural networks, we needed for our benchmark highdimensional data sets, which are labeled and have enough data points. ... Our experimental results in Table (3) show that, except for ISAAC, each method was able to achieve a quite similar VI. This indicates, that they were indeed able to find several non-redundant clusterings from the learned embeddings of the pretrained autoencoders.
Researcher Affiliation Academia 1Faculty of Computer Science, University of Vienna, Vienna, Austria 2Ludwig-Maximilians-Universit at M unchen, Munich, Germany 3MCML, 4ds:Uni Vie 1{lukas.miklautz, claudia.plant}@univie.ac.at 2{altinigneli, boehm, mautz}@dbs.ifi.lmu.de
Pseudocode No The paper describes its algorithms in text but does not include structured pseudocode or algorithm blocks.
Open Source Code Yes An implementation of ENRC and all experiments is available at https://gitlab.cs.univie.ac.at/lukas/enrcpublic.
Open Datasets Yes Concatenated-MNIST We extend the well known MNIST (Le Cun et al. 1998) data set by concatenating two digits side by side, resulting in 100 possible combinations, named CMNIST. ... GTSRB The German Traffic Sign Benchmark (GTSRB) data set (Houben et al. 2013)
Dataset Splits No The paper mentions training iterations and mini-batches, but does not provide specific details on train/validation/test dataset splits (e.g., percentages, sample counts, or explicit validation set usage).
Hardware Specification Yes We implemented ENRC in Python and trained our networks on a single NVIDIA RTX 2080 Ti. We ran the comparison method on a machine with four Intel(R) Xeon(R) CPU E5-2650 Cores and 32 GB RAM.
Software Dependencies No The paper mentions software like 'Python' and the 'scikit-learn package' but does not provide specific version numbers for these or other key software components.
Experiment Setup Yes For pretraining the autoencoder we use image augmentation (rotation, lighting, zooming), Adam (max-lr = 0.01, β1 = 0.9, β2 = 0.99) (Kingma and Ba 2014) with weight decay of 0.001. ... The discounting parameter α in Eq. 7, was set to 0.5 giving equal weight to new and old centers. The cluster reinitialization parameters were set to l = 10 and ns = 1, 000