On Finite-Sample Identifiability of Contrastive Learning-Based Nonlinear Independent Component Analysis

Authors: Qi Lyu, Xiao Fu

ICML 2022 | Conference PDF | Archive PDF | Plain Text | LLM Run Details

Reproducibility Variable Result LLM Response
Research Type Experimental Numerical experiments are used to validate the theorems. In this section, we validate our theoretical results using synthetic and real data experiments. Fig. 2 shows the n ICA performance in terms of MI using different network width R s under N= 5,000 and N= 10,000. Fig. 4 shows the averaged classification errors using SVM and logistic regression, respectively.
Researcher Affiliation Academia Qi Lyu 1 Xiao Fu 1 1School of EECS, Oregon State University, Corvallis, OR, United States. Correspondence to: Xiao Fu <xiao.fu@oregonstate.edu>, Qi Lyu <lyuqi@oregonstate.edu>.
Pseudocode No The paper does not contain any sections, figures, or blocks explicitly labeled as "Pseudocode" or "Algorithm".
Open Source Code No The paper does not contain any statements or links indicating that the source code for the described methodology is publicly available.
Open Datasets Yes In addition to synthetic data, we also use the EEG eye dataset from the UCI repository (Dua & Graff, 2017).
Dataset Splits No We use 12,000 data samples as the training set to learn h( ). Then, we train simple classifiers (i.e., SVM and logistic regression) using bs = h(x). The classifiers are tested using 3000 test samples. The paper specifies training and test sets, but no explicit validation set.
Hardware Specification No The paper does not specify any hardware details (e.g., GPU/CPU models, memory, or cloud computing instances) used for running the experiments.
Software Dependencies No For optimization, we use the Adam optimizer (Kingma & Ba, 2014) with an initial learning rate 5 10 4. We model h( ) and phi( ) using three-hidden-layer neural networks. The activation function is Re LU. we estimate the MI using kernel density estimation (Kozachenko & Leonenko, 1987). We compute the MI between each of the recovered by and the ground truth s s. Then, we use the Hungarian algorithm (Kuhn, 1955) to fix the permutation ambiguity. No specific version numbers for software are provided.
Experiment Setup Yes We model h( ) and phi( ) using three-hidden-layer neural networks. We test various R s (i.e., the number of hidden neurons) for each layer, where R {4, 8, 16, 32, 64, 128, 256, 512}. The activation function is Re LU. For optimization, we use the Adam optimizer (Kingma & Ba, 2014) with an initial learning rate 5 10 4.