Globally Convergent Variational Inference

Authors: Declan McNamara, Jackson Loper, Jeffrey Regier

NeurIPS 2024 | Conference PDF | Archive PDF | Plain Text | LLM Run Details

Reproducibility Variable Result LLM Response
Research Type Experimental In ablation studies and practical problems, we demonstrate that our results explain the behavior of NPE in non-asymptotic finite-neuron settings, and show that NPE outperforms ELBO-based optimization, which often converges to shallow local optima.
Researcher Affiliation Academia Declan Mc Namara Jackson Loper Jeffrey Regier Department of Statistics University of Michigan {declan, jaloper, regier}@umich.edu
Pseudocode No The paper does not contain any structured pseudocode or algorithm blocks.
Open Source Code Yes Our code is publicly available at https://github.com/declanmcnamara/gcvi_neurips.
Open Datasets Yes We use the MNIST dataset, freely available from torchvision under the BSD-3 License1
Dataset Splits No The paper uses generated data for several experiments and mentions using N=1000 MNIST digits, but does not provide explicit training, validation, or test dataset splits (percentages or counts) or refer to standard splits with citations.
Hardware Specification Yes We used Py Torch (Paszke et al., 2019) for our experiments in accordance with its license, and NVIDIA Ge Force RTX 2080 Ti GPUs.
Software Dependencies No We used Py Torch (Paszke et al., 2019) for our experiments in accordance with its license...
Experiment Setup Yes SGD was performed using the Adam optimizer with a learning rate of ρ = 0.0001. We employ a learning rate scheduler that scales the learning rate as O(1/I), where I denotes the number of iterations. All models were fitted for 200,000 stochastic gradient steps...