A Deep Semi-NMF Model for Learning Hidden Representations
Authors: George Trigeorgis, Konstantinos Bousmalis, Stefanos Zafeiriou, Bjoern Schuller
ICML 2014 | Conference PDF | Archive PDF | Plain Text | LLM Run Details
| Reproducibility Variable | Result | LLM Response |
|---|---|---|
| Research Type | Experimental | In order to evaluate this hypothesis, we have compared the performance of Deep Semi-NMF with that of other methods, on the task of clustering images of faces in 3 distinct datasets. |
| Researcher Affiliation | Academia | Department of Computing, Imperial College London, United Kingdom |
| Pseudocode | Yes | Algorithm 1 Suggested algorithm for training a Deep Semi-NMF model. |
| Open Source Code | Yes | Supplementary material including the implementation of Algorithm 1 and the proof of its convergence can be found at http://trigeorgis.com/deepseminmf. |
| Open Datasets | Yes | CMU Multi PIE: The first dataset we examine is The CMU Multi Pose, Illumination, and Expression (Multi PIE) Database (Gross et al., 2010)... CMU PIE: We also used a freely available version of CMU Pie (Sim et al., 2003)... XM2VTS: The Extended Multi Modal Verification for Teleservices and Security applications (XM2VTS) (Messer et al., 1999)... |
| Dataset Splits | No | The paper uses standard datasets but does not explicitly state specific train/validation/test splits, nor does it describe a cross-validation setup for these datasets. |
| Hardware Specification | No | No specific hardware details (e.g., GPU models, CPU types, or memory) used for the experiments are provided. |
| Software Dependencies | No | The paper mentions using NNDSVD, k-means clustering, and references specific algorithms, but does not provide specific version numbers for any software dependencies or libraries. |
| Experiment Setup | Yes | We specifically experimented with models that had a first hidden representation H1 with 625 features, and a second representation H2 with a number of features that ranged from 20 to 70. Again we chose these numbers after preliminary results showed us that the computational burden of additional features outweighed the performance increase obtained by increasing these numbers. We have set the maximum amount of iterations to 300 (usually 100 epochs are enough) and we use the convergence rule Ei 1 Ei κ max(1, Ei 1) in order to stop the process when reconstruction error (Ei) between the current and previous update is small enough. In our experiments we set κ = 10 6. |