Efficient Learning with a Family of Nonconvex Regularizers by Redistributing Nonconvexity

Authors: Quanming Yao, James Kwok

ICML 2016 | Conference PDF | Archive PDF | Plain Text | LLM Run Details

Reproducibility Variable Result LLM Response
Research Type Experimental In this section, we perform experiments on using the proposed procedure with (i) proximal algorithms (Sections 3.1 and 3.2); (ii) Frank-Wolfe algorithm (Section 3.3); and (iii) comparision with HONOR (Section 3.4).Results are shown in Table 2, As can be seen, all the nonconvex models obtain better RMSE and MABS than FISTA, and N2C is the fastest.
Researcher Affiliation Academia Quanming Yao QYAOAA@CSE.UST.HK James T. Kwok JAMESK@CSE.UST.HK Department of Computer Science and Engineering, Hong Kong University of Science and Technology, Hong Kong
Pseudocode Yes Algorithm 1 Frank-Wolfe algorithm for solving (7).
Open Source Code No The paper does not contain any explicit statement about releasing source code or provide a link to a code repository.
Open Datasets Yes We use the face data set JAFFE3, which contains 213 images with seven facial expressions: anger, disgust, fear, happy, neutral, sadness and surprise. Following (Liu & Ye, 2010), we resize each 256 256 image to 64 64. We also reuse their tree structure, which is based on pixel neighborhoods. 3http://www.kasrl.org/jaffe.htmland We use the Movie Lens data sets4 (Table 5), which have 4http://grouplens.org/datasets/movielens/and Experiments are performed on three large data sets, kdd2010a , kdd2010b and url 5 (Table 6). 5https://www.csie.ntu.edu.tw/ cjlin/ libsvmtools/datasets/binary.html
Dataset Splits Yes 50% of the data are used for training, another 25% for validation and the rest for testing. (Section 3.1) and 60% of the data are used for training, 20% for validation and the rest for testing. (Section 3.2) and Following (Yao et al., 2015), we use 50% of the ratings for training, 25% for validation and the rest for testing. (Section 3.3)
Hardware Specification Yes All algorithms are implemented in Matlab. The stopping criterion is reached when the relative change in objective is smaller than 10 8. Experiments are performed on a PC with Intel i7 CPU and 32GB memory.
Software Dependencies No All algorithms are implemented in Matlab. - No specific version of Matlab or any other software dependencies are provided.
Experiment Setup Yes We use a fixed stepsize of η = σmax(A A), while λ, µ in (12) are tuned by the validation set. For performance evaluation, we use the (i) testing root-mean-squared error (RMSE) on the predictions; (ii) mean absolute error of the obtained parameter ˆx with ground-truth x, MABS = ˆx x 1/10000; and (iii) CPU time. Each experiment is repeated 5 times, and the average performance reported. The stopping criterion is reached when the relative change in objective is smaller than 10 8.