Unsupervised Metric Learning with Synthetic Examples
Authors: Ujjal Kr Dutta, Mehrtash Harandi, C. Chandra Sekhar3834-3841
AAAI 2020 | Conference PDF | Archive PDF | Plain Text | LLM Run Details
| Reproducibility Variable | Result | LLM Response |
|---|---|---|
| Research Type | Experimental | Classification on benchmark datasets We first compare the proposed SUML approach against the following recently proposed State-Of-The-Art (SOTA) supervised DML approaches that make use of class labels: JDRML (Harandi, Salzmann, and Hartley 2017), LRGMML (Bhutani et al. 2018), AML (Chen et al. 2018) and MDMLCLDD (Xie et al. 2018). We use the following benchmark datasets: AT&T Faces(Samaria and Harter 1994), COIL20 (Nene et al. 1996), Isolet (Lichman 2013) and USPS (Hull 1994). For each dataset, we perform 10 random splits with 70%-30% train-test ratio, and report the standard deviation along with the classification accuracies on the test data. [...] As seen in Table 1, our method performs competitive despite being unsupervised. We also compare our method against unsupervised manifold learning techniques NPE (He et al. 2005) and LPP (He and Niyogi 2003), which easily get outperformed by SUML (Table 2). |
| Researcher Affiliation | Academia | Ujjal Kr Dutta, ,1 Mehrtash Harandi,*,2 C Chandra Sekhar ,3 Dept. of Computer Science and Eng., Indian Institute of Technology Madras, India *Dept. of Electrical and Computer Systems Eng., Monash University, Australia 1ukd@cse.iitm.ac.in, 2mehrtash.harandi@monash.edu, 3chandra@cse.iitm.ac.in |
| Pseudocode | Yes | Algorithm 1 stochastic SUML (s SUML) |
| Open Source Code | No | The paper does not contain any explicit statements about making its code open source or providing a repository link. |
| Open Datasets | Yes | We use the following benchmark datasets: AT&T Faces(Samaria and Harter 1994), COIL20 (Nene et al. 1996), Isolet (Lichman 2013) and USPS (Hull 1994). [...] For the task of Zero-Shot Learning (ZSL) we used two action recognition datasets: JHMDB (Jhuang et al. 2013) [...] and HMDB (Kuehne et al. 2011) [...]. For DML in presence of noise, we used the Aw A2 (Xian et al. 2018) and Fashion-MNIST (Xiao, Rasul, and Vollgraf 2017) datasets. [...] original MNIST dataset (Le Cun et al. 1998). [...] STL-10 (Coates, Ng, and Lee 2011) and Image Net (Russakovsky et al. 2015). [...] CUB 200 (Welinder et al. 2010), Cars 196 (Krause et al. 2013) and Stanford Online Products (SOP) (Oh Song et al. 2016). |
| Dataset Splits | No | The paper specifies a "70%-30% train-test ratio" and mentions hyperparameter tuning for 'alpha' and 'lambda', but it does not explicitly state the use of a separate validation dataset or a specific cross-validation strategy for hyperparameter tuning. The term 'validation' is only used once in the context of a dataset name for evaluation but not as a split in their own experiments. |
| Hardware Specification | No | The paper mentions "runtime" and "computational time complexity" but does not provide any specific details about the hardware (e.g., CPU, GPU models, memory) used for running the experiments. |
| Software Dependencies | No | The paper mentions using "Manopt toolbox (Boumal et al. 2014)" and "Mat Conv Net (Vedaldi and Lenc 2015) tool" but does not specify version numbers for these software components or any other libraries. |
| Experiment Setup | Yes | α > 0 can be intuitively seen as an angle with respect to a triplet (Wang et al. 2017). We do not tune it, but set it to 40 for all our experiments. [...] We do not give too high weightage to the regularizer term, and hence set λ to a low value of 0.5 for all our experiments. [...] trained for a maximum of num ep = 30 epochs for all datasets. [...] For the Image Net dataset, we followed the Im Net-2 protocol as in (Kodirov, Xiang, and Gong 2017). In both STL-10 and Image Net, we fix maxiter = 10, α = 40 , λ = 0.5 in Algorithm 1. Mini-batch size is 120, and embedding size is 64. [...] we set the embedding size to 512, except for Cars, where we set it as 128. We fix α = 45 and λ = 0.5. We used mini-batch size of 120 and set maxiter = 10 in Algorithm 1. |