Deep Adaptive Design: Amortizing Sequential Bayesian Experimental Design

Authors: Adam Foster, Desi R Ivanova, Ilyas Malik, Tom Rainforth

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

Reproducibility Variable Result LLM Response
Research Type Experimental We apply DAD to a range of problems relevant to applications such as epidemiology, physics and psychology. We find that DAD is able to accurately amortize experiments, opening the door to running adaptive BOED in real time.
Researcher Affiliation Academia 1Department of Statistics, University of Oxford, UK 2Work undertaken whilst at the University of Oxford.
Pseudocode Yes Algorithm 1 Deep Adaptive Design (DAD) Input: Prior p(θ), likelihood p(y|θ, ξ), number of steps T Output: Design network πφ
Open Source Code Yes Code is publicly available at https://github.com/ae-foster/dad.
Open Datasets No The paper describes generating training examples (e.g., "We generate 200,000 training examples") rather than using a publicly available, pre-existing dataset with specific access information (URL, citation with authors/year, or repository).
Dataset Splits No We do not use an explicit validation set during training.
Hardware Specification Yes All experiments were run on a 2.3 GHz 8-Core Intel Core i9 processor with 16GB of DDR4 memory. For the location finding and hyperbolic discounting experiments, the training was conducted on a single NVIDIA Quadro RTX 4000 GPU.
Software Dependencies No We implement DAD by extending Py Torch (Paszke et al., 2019) and Pyro (Bingham et al., 2018). While these are mentioned, specific version numbers are not provided.
Experiment Setup Yes We use the Adam optimizer (Kingma & Ba, 2014) with a learning rate of 0.001. For the training of DAD we used L=30 contrastive samples... We generate 200,000 training examples (batches of 128) for the location finding problem and 50,000 for the hyperbolic discounting problem.