Benchmarking Deep Inverse Models over time, and the Neural-Adjoint method

Authors: Simiao Ren, Willie Padilla, Jordan Malof

NeurIPS 2020 | Conference PDF | Archive PDF | Plain Text | LLM Run Details

Reproducibility Variable Result LLM Response
Research Type Experimental Using this metric, we compare several state-of-the-art inverse modeling approaches on four benchmark tasks: two existing tasks, a new 2-dimensional sinusoid task, and a challenging modern task of meta-material design.
Researcher Affiliation Academia Dept. of Electrical and Computer Engineering Duke University Durham, NC 27705
Pseudocode No The paper does not contain any pseudocode or clearly labeled algorithm blocks.
Open Source Code Yes We release code for all inverse models, as well as (fast) simulation software for each benchmark problem, so that other researchers can easily repeat our experiments. 1https://github.com/Benson Ren/BDIMNNA
Open Datasets Yes Inspired by the recent benchmark study [2], we include two popular existing tasks: ballistics targeting (D1), and robotic arm control (D3). For these two tasks we use the same experimental designs as [2], including their simulator (i.e., forward model) parameters, simulator sampling procedures, and their training/testing splits.
Dataset Splits Yes For these two tasks we use the same experimental designs as [2], including their simulator (i.e., forward model) parameters, simulator sampling procedures, and their training/testing splits. All details can be found in [2] and our supplement.
Hardware Specification No The paper mentions 'common hardware' but does not provide specific details like GPU or CPU models.
Software Dependencies No The paper mentions using 'modern deep learning software packages' but does not list specific software names with version numbers.
Experiment Setup No All models utilized the same training and testing data, batch size, and stopping criteria (for training). In those cases where model hyperparameters were not available from [2], we budgeted approximately one day of computation time (on common hardware) to optimize hyperparameters, while again constraining model sizes. Full implementation details can be found in the supplementary material.