$\texttt{NAISR}$: A 3D Neural Additive Model for Interpretable Shape Representation

Authors: Yining Jiao, Carlton Jude ZDANSKI, Julia S Kimbell, Andrew Prince, Cameron P Worden, Samuel Kirse, Christopher Rutter, Benjamin Shields, William Alexander Dunn, Jisan Mahmud, Marc Niethammer

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

Reproducibility Variable Result LLM Response
Research Type Experimental We evaluate NAISR with respect to shape reconstruction, shape disentanglement, shape evolution, and shape transfer on three datasets: 1) Starman, a simulated 2D shape dataset; 2) the ADNI hippocampus 3D shape dataset; and 3) a pediatric airway 3D shape dataset. Our experiments demonstrate that NAISR achieves excellent shape reconstruction performance while retaining interpretability.
Researcher Affiliation Academia 1University of North Carolina at Chapel Hill, 2Wake Forest School of Medicine, 3The Ohio State University College of Medicine
Pseudocode No The paper includes mathematical formulations and figures but no explicit pseudocode or algorithm blocks.
Open Source Code Yes Our code is publicly available. ... We have submitted our source code. The implementation of NAISR will be made publicly available.
Open Datasets Yes Data used in preparation of this article were obtained from the Alzheimer s Disease Neuroimaging Initiative (ADNI) database (adni.loni.usc.edu). ... Starman, a simulated 2D shape dataset (Bône et al., 2020).
Dataset Splits No The paper mentions "80%-20% train-test split" for datasets but does not explicitly describe a separate validation split or how validation data was partitioned or used.
Hardware Specification Yes The model is trained on an Nvidia Ge Force RTX 3090 GPU for approximately 12 hours for 3000 epochs for the airway dataset.
Software Dependencies No The paper mentions using "Adam (Kingma & Ba, 2014)" and "SIREN (Sitzmann et al., 2020)" but does not specify version numbers for these or other software components.
Experiment Setup Yes There are 256 hidden units in each layer. ... We train NAISR for 3000 epochs for the airway dataset and 300 epochs for the ADNI hippocampus and Starman datasets using Adam (Kingma & Ba, 2014) with a learning rate 5e 5 and batch size of 64. ... During training, λ1 = λ5 = 1 10; λ2 = 3 10; λ3 = 1 10, λ4 = 1 102. For Llat, λ6 = 2 L; σ = 0.01 ... During inference, the latent codes are optimized for Nt iterations with a learning rate of 5e 3. Nt is set to 800 for the pediatric airway dataset; Nt is set to 200 for the Starman and ADNI Hippocampus datasets.