Notice: The reproducibility variables underlying each score are classified using an automated LLM-based pipeline, validated against a manually labeled dataset. LLM-based classification introduces uncertainty and potential bias; scores should be interpreted as estimates. Full accuracy metrics and methodology are described in [1].

Distilling Dataset into Neural Field

Authors: Donghyeok Shin, HeeSun Bae, Gyuwon Sim, Wanmo Kang, Il-chul Moon

ICLR 2025 | Venue PDF | LLM Run Details

Reproducibility Variable Result LLM Response
Research Type Experimental Through extensive experiments, we demonstrate that DDi F achieves superior performance on several benchmark datasets, extending beyond the image domain to include video, audio, and 3D voxel. We release the code at https://github.com/aailab-kaist/DDi F.
Researcher Affiliation Academia 1Korea Advanced Institute of Science and Technology (KAIST), 2summary.ai EMAIL
Pseudocode Yes Algorithms 1 and 2 specify a training procedure and decoding process of DDi F, respectively.
Open Source Code Yes We release the code at https://github.com/aailab-kaist/DDi F.
Open Datasets Yes Image Net-Subset (Howard, 2019; Cazenavette et al., 2022), CIFAR-10 (Krizhevsky et al., 2009), CIFAR-100 (Krizhevsky et al., 2009), Mini UCF (Wang et al., 2024; Khurram, 2012), Mini Speech Commands (Kim et al., 2022; Warden, 2018), Model Net (Wu et al., 2015), Shape Net (Chang et al., 2015)
Dataset Splits Yes Each class contains 5,000 images for training and 1,000 images for testing. ... Each class is split into 500 for training and 100 for testing. ... The dataset consists of 8 classes, and each class has 875/125 data for training/testing, respectively.
Hardware Specification Yes We use a mixture of RTX 3090, L40S, and Tesla A100 to run our experiments.
Software Dependencies No The paper mentions software components like 'Adam optimizer' and 'Kornia' but does not specify their version numbers. For example, 'We utilize Adam optimizer (Kingma & Ba, 2017)' and 'We adopt ZCA whitening ... with the Kornia (Riba et al., 2020) implementation'.
Experiment Setup Yes We fix the iteration number and learning rate for warm-up initialization of synthetic neural field as 5,000 and 0.0005. ... We run 15,000 iterations for TM and 20,000 iterations for DM. ... We provide the detailed hyperparameters in Table 9.