Manifold Diffusion Fields

Authors: Ahmed A. A. Elhag, Yuyang Wang, Joshua M. Susskind, Miguel Ángel Bautista

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

Reproducibility Variable Result LLM Response
Research Type Experimental Empirical results on multiple datasets and manifolds including challenging scientific problems like weather prediction or molecular conformation show that MDF can capture distributions of such functions with better diversity and fidelity than previous approaches.
Researcher Affiliation Industry Ahmed A. Elhag , Yuyang Wang, Joshua M. Susskind, Miguel Angel Bautista Apple {aa_elhag, yuyang_wang4, jsusskind, mbautistamartin}@apple.com
Pseudocode Yes Algorithm 1 Training; Algorithm 2 Sampling
Open Source Code No The paper mentions using a 'modified version of the publicly available repository is used for Perceiver IO 2.' and an 'ERA5 dataset... is available at' external links, but does not explicitly state that the code for MDF itself is open-sourced or provide a specific link for it.
Open Datasets Yes MNIST (Le Cun et al., 1998) and Celeb A-HQ (Karras et al., 2018) datasets, where images are texture mapped into the meshes using (Sullivan & Kaszynski, 2019), models are evaluated on the standard tests sets for these datasets. Following the standard setting for molecule conformer prediction we use the GEOM-QM9 dataset (Ruddigkeit et al., 2012; Ramakrishnan et al., 2014) which contains 130K molecules ranging from 10 to 40 atoms.
Dataset Splits Yes We use the same train/val/test splits as Torsional Diffusion (Jing et al., 2022) and use the same metrics to compare the generated and ground truth conformer ensembles: Average Minimum RMSD (AMR) and Coverage.
Hardware Specification Yes Each model was trained on an machine with 8 Nvidia A100 GPUs, we trained models for 3 days.
Software Dependencies No The paper mentions 'Perceiver IO' and 'Adam' optimizer, but does not provide specific version numbers for key software dependencies like Python, PyTorch, or CUDA.
Experiment Setup Yes Table 8: Hyperparameters and settings for MDF on different manifolds. (train res., #context set, #query set, #eigenfuncs (k), #freq pos. embed t, #latents, #dim latents, #blocks, #dec blocks, #self attends per block, #self attention heads, #cross attention heads, batch size, lr, epochs)