Mixture weights optimisation for Alpha-Divergence Variational Inference

Authors: Kamélia Daudel, randal douc

NeurIPS 2021 | Conference PDF | Archive PDF | Plain Text | LLM Run Details

Reproducibility Variable Result LLM Response
Research Type Experimental Lastly, we compare numerically the behavior of the unbiased Power Descent and of the biased Rényi Descent and we discuss the potential advantages of one algorithm over the other. ... Finally, we run some numerical experiments in Section 5 to compare the behavior of the Power Descent and of the Rényi Descent altogether, before discussing the potential benefits of one approach over the other. ... 5 Simulation study
Researcher Affiliation Academia 1: LTCI, Télécom Paris, Institut Polytechnique de Paris, France 2: Department of Statistics, University of Oxford, United Kingdom 3: SAMOVAR, Télécom Sud Paris, Institut Polytechnique de Paris, France
Pseudocode Yes Algorithm 1: Power descent one-step transition (Γ(v) = [(α 1)v + 1]η/(1 α)) ... Algorithm 2: Complete Exploitation-Exploration Algorithm
Open Source Code Yes The code for all the subsequent numerical experiments is available at https://github.com/ kdaudel/Mixture Weights Alpha VI.
Open Datasets No The paper uses a synthetic target distribution defined as "p(y) = c [0.5N(y; sud, Id) + 0.5N(y; sud, Id)]" for its simulation study, rather than a pre-existing, publicly available dataset that would require a public link or citation. Therefore, there is no information about publicly available training data.
Dataset Splits No The paper does not describe specific training, validation, or test dataset splits. The experiments are conducted on a synthetically defined target density, and the performance is assessed through metrics like the Variational Rényi bound, not through standard dataset splits.
Hardware Specification No The paper does not provide any specific details about the hardware used to run the experiments (e.g., CPU/GPU models, memory, or cloud computing specifications). It only mentions running numerical experiments.
Software Dependencies No The paper mentions that the code is available on GitHub and written in Python, but it does not specify any particular software libraries, frameworks, or their version numbers (e.g., PyTorch 1.9, TensorFlow 2.x, or specific Python library versions) that would be necessary to replicate the experiments.
Experiment Setup Yes We take J = 100, M {100, 1000, 2000}, α = 0.5, κ = 0, η0 = 0.3 and q0 is a centered normal distribution with covariance matrix 5Id. We let T = 10, N = 20 and we replicate the experiment 100 times independently for each algorithm.