Efficiently testing local optimality and escaping saddles for ReLU networks

Authors: Chulhee Yun, Suvrit Sra, Ali Jadbabaie

ICLR 2019 | Conference PDF | Archive PDF | Plain Text | LLM Run Details

Reproducibility Variable Result LLM Response
Research Type Experimental For experiments, we used artificial datasets sampled iid from standard normal distribution, and trained 1-hidden-layer Re LU networks with squared error loss. ... Table 1 summarizes the results.
Researcher Affiliation Academia Chulhee Yun, Suvrit Sra & Ali Jadbabaie Massachusetts Institute of Technology Cambridge, MA 02139, USA {chulheey,suvrit,jadbabai}@mit.edu
Pseudocode Yes Algorithm 1 SOSP-CHECK (Rough pseudocode); Algorithm 2 SOSP-CHECK; Algorithm 3 FO-SUBDIFF-ZERO-TEST; Algorithm 4 FO-INCREASING-TEST; Algorithm 5 SO-TEST
Open Source Code No The paper does not provide any statements or links indicating that source code for the described methodology is publicly available.
Open Datasets Yes For experiments, we used artificial datasets sampled iid from standard normal distribution
Dataset Splits No The paper mentions using artificial datasets and training but does not specify train, validation, or test splits by percentages, counts, or by referencing standard splits.
Hardware Specification No The paper describes the experimental setup but does not provide any specific hardware details such as GPU/CPU models or memory.
Software Dependencies No The paper mentions using Adam (Kingma & Ba, 2014) and L-BFGS-B (Byrd et al., 1995) but does not provide specific version numbers for these or any other software dependencies.
Experiment Setup Yes we ran Adam (Kingma & Ba, 2014) using full-batch (exact) gradient for 200,000 iterations and decaying step size (start with 10 3, 0.2 decay every 20,000 iterations).