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].
Determinant Estimation under Memory Constraints and Neural Scaling Laws
Authors: Siavash Ameli, Chris Van Der Heide, Liam Hodgkinson, Fred Roosta, Michael W. Mahoney
ICML 2025 | Venue PDF | LLM Run Details
| Reproducibility Variable | Result | LLM Response |
|---|---|---|
| Research Type | Experimental | Numerical experiments are presented in Section 4, and we conclude in Section 5. |
| Researcher Affiliation | Academia | 1Department of Statistics, University of California, Berkeley CA, USA 2International Computer Science Institute, Berkeley CA, USA 3Dept. of Electrical and Electronic Engineering, University of Melbourne, Australia 4School of Mathematics and Statistics, University of Melbourne, Australia 5CIRES and School of Mathematics and Physics, University of Queensland, Australia 6Lawrence Berkeley National Laboratory, Berkeley CA, USA. Correspondence to: Chris van der Heide <EMAIL>. |
| Pseudocode | Yes | A pseudo code and further efficient implementation details of the presented method can be found as Algorithm D.1. |
| Open Source Code | Yes | we provide a high-performance Python package detkit, which implements the presented algorithms and can be used to reproduce the results of this paper. (Further details with URLs are in Appendix I: detkit is available for installation from Py PI (https://pypi.org/project/detkit), the documentation can be found at https://ameli.github.io/detkit, and the source code is available at https://github.com/ameli/detkit.) |
| Open Datasets | Yes | Res Net9, Res Net18, and Res Net50 (He et al., 2016) trained on the CIFAR-10 dataset (Krizhevsky, 2009), and Mobile Net (Howard et al., 2017) trained on the MNIST dataset (Le Cun et al., 1998). |
| Dataset Splits | No | a Res Net50 model was trained on a subset of 1000 datapoints from the CIFAR-10 dataset with d = 10 classes. The paper mentions using "subsets" of datasets for experiments (e.g., "1000 datapoints from the CIFAR-10 dataset") but does not provide specific train/test/validation splits or methodology for partitioning the data for their experiments. |
| Hardware Specification | Yes | Costs and wall time are based on an NVIDIA H100 GPU ($2/hour) and an 8core 3.6GHz CPU ($0.2/hour) using Amazon pricing" (Table 1 caption). Also, "Experiments in this section were conducted on a desktop-class device with an AMD Ryzen 7 5800X processor, NVIDIA RTX 3080, and 64GB RAM." (Section 4.1). Further, "This computation was carried out on an NVIDIA Grace Hopper GH200 GPU over 244 hours for Res Net50." (Section 4.2). |
| Software Dependencies | No | All neural networks (e.g., Res Net9, Res Net50) were trained using 32-bit precision, which is the default and standard practice in most deep learning frameworks such as Py Torch." The paper mentions PyTorch but does not specify a version number. Other software like numpy, zarr, and imate are used or implemented but their specific versions are not provided. |
| Experiment Setup | Yes | FLODANCE with q = 6, n0 = 100, and ns = 5000 achieves an absolute error of just 0.2% for ˆℓ50,000 on Res Net9... Similarly, for Res Net50, FLODANCE with q = 4, n0 = 100, and ns = 5000 achieves an absolute error of 0.02% with the same speedup." (Section 4.2) The paper also specifies neural network models (ResNet9, ResNet18, ResNet50, MobileNet), datasets (CIFAR-10, MNIST), and floating-point precisions (16-, 32-, and 64-bit) used for comparisons. |