Interaction-Force Transport Gradient Flows

Authors: Egor Gladin, Pavel Dvurechenskii, Alexander Mielke, Jia-Jie Zhu

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

Reproducibility Variable Result LLM Response
Research Type Experimental We then empirically demonstrate the use of the IFT gradient flow for the MMD inference task. Compared to the original MMD-energy-flow algorithm of Arbel et al. [2019], IFT flow does not suffer issues such as the collapsing-to-mode issue. Leveraging the first-principled spherical IFT gradient flow, our method does not require a heuristic noise injection that is commonly tuned over the iterations in practice; see [Korba et al., 2021] for a discussion.Our method can also be viewed as addressing a long-standing issue of the kernel-mean embedding methods [Smola et al., 2007, Muandet et al., 2017, Lacoste-Julien et al., 2015] for optimizing the support of distributions.4 Numerical Example The overall goal of the numerical experiments is to approximate the target measure π by minimizing the squared MMD energy, i.e., min µ A P MMD2(µ, π).
Researcher Affiliation Academia Egor Gladin Humboldt University of Berlin Berlin, Germany & HSE University egorgladin@yandex.ru Pavel Dvurechensky Weierstrass Institute for Applied Analysis and Stochastics Berlin, Germany pavel.dvurechensky@wias-berlin.de Alexander Mielke Humboldt University of Berlin & WIAS Berlin, Germany alexander.mielke@wias-berlin.de Jia-Jie Zhu Weierstrass Institute for Applied Analysis and Stochastics Berlin, Germany jia-jie.zhu@wias-berlin.de
Pseudocode Yes We summarize the resulting overall IFT particle gradient descent from the JKO splitting scheme in Algorithm 1 in the appendix. ... Algorithm 1 A JKO-splitting for IFT particle gradient descent
Open Source Code Yes We provide the code for the implementation at https://github.com/egorgladin/ift_flow.
Open Datasets Yes The overall goal of the numerical experiments is to approximate the target measure π by minimizing the squared MMD energy, i.e., min µ A P MMD2(µ, π). In all the experiments, we have access to the target measure π in the form of samples yi π. This setting was studied in [Arbel et al., 2019] as well as in many deep generative model applications. ... µ0 N(5 1, I) and π N 0, 1 1/2 1/2 2 . ... this time the target is a mixture of equally weighted Gaussian distributions, N 0, 1 1/2 1/2 2 , N 3 1 , I , N 1 4 , 3 1/2 1/2 1 .
Dataset Splits No The paper does not explicitly mention a 'validation set' or a 'validation split' of the data. The experiments focus on approximating a target distribution and sampling, rather than supervised learning with distinct validation data.
Hardware Specification No The main text of the paper does not provide specific hardware details such as GPU/CPU models, processor types, or memory amounts used for running the experiments. The NeurIPS checklist states that 'Our experiments are small-scale and can be reproduced on a standard laptop', but this general statement lacks the specificity required for hardware reproduction.
Software Dependencies No The paper does not explicitly provide specific software dependencies with version numbers (e.g., library names like PyTorch with a version number) within its main text or appendix.
Experiment Setup Yes A Gaussian kernel with bandwidth σ = 10 was used. For all three algorithms, we chose the largest stepsize that didn t cause unstable behavior, τ = 50. The parameter η in (23) was set to 0.1.