Sparse Subspace Clustering by Learning Approximation _0 Codes
Authors: Jun Li, Yu Kong, Yun Fu
AAAI 2017 | Conference PDF | Archive PDF | Plain Text | LLM Run Details
| Reproducibility Variable | Result | LLM Response |
|---|---|---|
| Research Type | Experimental | Experiments on handwritten digit and face clustering show that our method not only quickly computes the ℓ0 code, but also outperforms the relevant subspace clustering methods in clustering results. In particular, our method achieves the state-of-the-art clustering accuracy (94.32%) on MNIST. In this section, we first introduce the experimental settings. Second, we verify that our method (SSC-LOMP) has better subspace preserving property than SSC-OMP, and achieves the state-of-the-art clustering results on MNIST dataset. |
| Researcher Affiliation | Academia | 1Department of Electrical and Computer Engineering, Northeastern University, Boston, MA, 02115, USA. 2College of Computer and Information Science, Northeastern University, Boston, MA, 02115, USA. junl.mldl@gmail.com, {yukong,yunfu}@ece.neu.edu |
| Pseudocode | Yes | Algorithm 1 Orthogonal Matching Pursuit (OMP) Algorithm 2 Learning W1 and W2 by the gradient descent (GD) algorithm Algorithm 3 Learned OMP (LOMP) algorithm Algorithm 4 Sparse Subspace Clustering by the learned OMP Algorithm (SSC-LOMP) |
| Open Source Code | No | The paper states: "The codes provided by the respective authors are used to calculate the representation matrix X in all experiments, except for SSC-LISTA, which is coded by us, because it is first to apply into subspace clustering." This indicates use of other authors' code and their own implementation of SSC-LISTA, but does not provide concrete access to the source code for the SSC-LOMP method described in the paper. |
| Open Datasets | Yes | Datasets. MNIST1 has 70,000 examples with 28 28 pixel greyscale images of handwritten digits 0-9. We use a scattering convolution network (Bruna and Mallat 2013) to compute the feature vectors of each image by following the experimental data settings (You, Robinson, and Vidal 2016). Each feature vector is of length size 3,472 and the features are reduced to 500 dimensions by PCA. We randomly choose {50, 100, 200, 400, 600} images from each class for all experiments. 1http://yann.lecun.com/exdb/mnist/ Extended Yale B2 contains frontal face images of 38 individuals with 64 different illumination conditions. The size of each image is 192 168, and the original face images are downsampled to 48 42 pixels. We randomly pick ni {2, 10, 20, 30, 38} classes, and take all the images of them as the data to be clustered. 2http://vision.ucsd.edu/~leekc/Ext_Yale_Database/Ext_Yale_B.html |
| Dataset Splits | No | The paper does not provide specific dataset split information (exact percentages, sample counts, or detailed splitting methodology) for training, validation, or testing. It mentions randomly choosing images/classes for experiments and varying the total number of points/classes, but no explicit train/validation/test splits are detailed. |
| Hardware Specification | No | The paper states: "all algorithms run on Matlab 2015 and Windows 7." This specifies software and operating system, but does not provide any specific hardware details such as CPU model, GPU model, or memory capacity used for running the experiments. |
| Software Dependencies | Yes | All algorithms run on Matlab 2015 and Windows 7. |
| Experiment Setup | Yes | In the proposed method (SSC-LOMP), there are three important parameters θ, m, and ε, which control the threshold value of shrinkage function, the the number of hidden units and learning rate. Actually, θ, m and ε are the typical parameters in neural networks (Gregor and Le Cun 2010; Li, Chang, and Yang 2015). So, they are easy to choose in our given sets: θ = {0.2, 0.1, 0.05, 0.03, 0.02, 0.01, 0.005, 0.001, 0.0001}, M = {500, 1000, 1200, 2000, 2500, 3000, 3500, 4000, 5500} and ε = {0.01, 0.001, 0.0001, 0.00001}. In all experiments, kmax is set to 5 (Extended Yale B) or 10 (MNIST), the regularization parameter of weights β is set to 0.00001, shrinkage and tanh (or Re LU) are voted as the activation function f( ) and g( ), and the number of training iterations tmax is less than 10. |