Annealed Flow Transport Monte Carlo

Authors: Michael Arbel, Alex Matthews, Arnaud Doucet

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

Reproducibility Variable Result LLM Response
Research Type Experimental We demonstrate experimentally the benefits and limitations of our methodology on a variety of applications.
Researcher Affiliation Collaboration 1Gatsby Computational Neuroscience Unit, University College London 2Deep Mind. Correspondence to: Michael Arbel <michael.n.arbel@gmail.com>, Alexander G. D. G. Matthews <alexmatthews@google.com>, Arnaud Doucet <arnauddoucet@google.com>.
Pseudocode Yes Algorithm 1 Annealed Flow Transport
Open Source Code No We plan to make the code available within https://github.com/deepmind.
Open Datasets Yes For our next example, we trained a variational autoencoder (Kingma and Welling, 2014; Rezende et al., 2014) with convolution on the binarized MNIST dataset (Salakhutdinov and Murray, 2008)
Dataset Splits No As discussed in Section 3.3, we use three sets of particles—train, test and validation which improves robustness, avoids overfitting the flow to the particles and gives unbiased estimates of Z when using the test set. The paper states that validation sets are used but does not provide specific split percentages or counts for reproducibility.
Hardware Specification No The paper does not provide specific hardware details such as GPU models, CPU types, or cloud instance specifications used for running the experiments. It only mentions "modern hardware" in general terms.
Software Dependencies No The paper mentions various software components and libraries (e.g., JAX, Haiku, Optax, TensorFlow Distributions, Adam) but does not provide specific version numbers for these dependencies, which are necessary for reproducibility.
Experiment Setup Yes We tune the step size to have a reasonable acceptance probability based on preliminary runs of SMC using a modest K. Then for larger K experiments, we linearly interpolate the step sizes chosen on the preliminary runs. We always use a linearly spaced geometric schedule and the initial distribution is always a multivariate standard normal. We repeat experiments 100 times. For the AFT flow we used an affine transformation with diagonal linear transformation matrix.