Conditioning by adaptive sampling for robust design

Authors: David Brookes, Hahnbeom Park, Jennifer Listgarten

ICML 2019 | Conference PDF | Archive PDF | Plain Text | LLM Run Details

Reproducibility Variable Result LLM Response
Research Type Experimental We perform two main sets of experiments. In the first, we use simple, one-dimensional, toy examples to demonstrate some of the main points about our approach. In the second set of experiments, we ground ourselves in a real protein fluorescence data set, conducting extensive simulations to compare methods.
Researcher Affiliation Academia 1Biophysics Graduate Group, UC Berkeley, CA 2Department of Biochemistry, University of Washington, Seattle, WA 3Institute for Protein Design, University of Washington, Seattle, WA 4EECS Department, UC Berkeley, CA.
Pseudocode Yes In Algorithm 1 in the Supplemental Information, we outline our complete procedure when the prior and generative model are both latent variable models of the same parametric form (e. g., both a VAE).
Open Source Code No The paper does not provide a direct link to a code repository or an explicit statement about releasing the source code for the described methodology.
Open Datasets Yes We anchored our experiments on a real protein fluorescence data set (Sarkisyan et al., 2016) (see Supplementary Information).
Dataset Splits No The paper mentions using 5,000 samples for training oracles and refers to a 'hold-out set' and 'test set', but does not provide specific percentages, sample counts, or detailed methodology for train/validation/test dataset splits.
Hardware Specification No The paper does not provide specific hardware details such as GPU/CPU models, processor types, or memory amounts used for running experiments.
Software Dependencies No The paper mentions various models and methods but does not provide specific software dependency details, such as library names with version numbers, needed for replication.
Experiment Setup Yes We use a search distribution that is the same parametric form as the prior, that is, a Gaussian distribution and run our method for 50 iterations, with the quantile update parameter, Q = 1 (meaning that γ(t) will be set to the maximum over the sampled mean oracle values at iteration t) and M = 100 samples taken at each iteration.