Geometric Analysis of Nonconvex Optimization Landscapes for Overcomplete Learning

Authors: Qing Qu, Yuexiang Zhai, Xiao Li, Yuqian Zhang, Zhihui Zhu

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

Reproducibility Variable Result LLM Response
Research Type Experimental Finally, our theoretical results are justified by numerical simulations.
Researcher Affiliation Academia Qing Qu Center for Data Science New York University qq213@nyu.edu Yuexiang Zhai EECS UC Berkeley ysz@berkeley.edu Xiao Li Eletronic Engineering CUHK xli@ee.cuhk.edu.hk Yuqian Zhang Electrical & Computer Engineering Rutgers University yqz.zhang@rutgers.edu Zhihui Zhu Electrical & Computer Engineering University of Denver zhihui.zhu@du.edu
Pseudocode Yes Algorithm 1 Finding one filter with data-driven initialization
Open Source Code No The paper provides a link to its arXiv preprint ('https://arxiv.org/abs/1912.02427') which contains the full version of the paper, but there is no explicit statement or link for open-source code for the methodology described.
Open Datasets No The paper states: 'We generate data Y AX, with dictionary A P Rnˆm being UNTF, and sparse code X P Rmˆp i.i.d. BGpθq.' and 'for CDL, we generate measurement according to Equation (1.2) with K 3, where the filters ta0ku K k 1 are drawn uniformly from the sphere Sn 1, and xik i.i.d. BGpθq.' This indicates generated synthetic data, not a publicly available or open dataset.
Dataset Splits No The paper generates its own data for simulations and does not describe explicit train/validation/test splits of a dataset. It focuses on parameters used for data generation (e.g., n=3, m=4, p=2*10^4).
Hardware Specification No The paper does not provide any specific details about the hardware (e.g., GPU/CPU models, memory) used for running the numerical simulations.
Software Dependencies No The paper does not list any specific software dependencies with version numbers (e.g., Python, PyTorch, specific libraries or solvers) used for implementing the methods or running experiments.
Experiment Setup Yes We generate data Y AX, with dictionary A P Rnˆm being UNTF, and sparse code X P Rmˆp i.i.d. BGpθq. To judge the success recovery of one column of A, let ρe min 1ďiďm p1 |xq , ai{ }ai}y|q . We have ρe 0 when q PSn 1paiq, thus we assume a recovery is successful if ρe ă 5 ˆ 10 2. ... First, we fix θ 0.1, and test the limit of the overcompleteness K m{n we can achieve by plotting the phase transition on pm, nq in log scale. For each pair of pm, nq, we repeat the experiment for 12 times. ... Parameters: n 64, θ 0.1, K 3, p 1 ˆ 104.