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].

Compositional Curvature Bounds for Deep Neural Networks

Authors: Taha Entesari, Sina Sharifi, Mahyar Fazlyab

ICML 2024 | Venue PDF | LLM Run Details

Reproducibility Variable Result LLM Response
Research Type Experimental Finally, we demonstrate the efficacy of our method on classification tasks using the MNIST and CIFAR-10 datasets.
Researcher Affiliation Academia Taha Entesari * 1 Sina Sharifi * 1 Mahyar Fazlyab 1 1Department of Electrical and Computer Engineering, Johns Hopkins University, Baltimore, United States of America. Correspondence to: Mahyar Fazlyab <EMAIL>.
Pseudocode Yes Algorithm 1 Compositional Curvature Estimation of Neural Networks
Open Source Code Yes Our code is available at https://github.com/o4lc/Compositional Curvature-Bounds-for-DNNs.
Open Datasets Yes Finally, we demonstrate the efficacy of our method on classification tasks using the MNIST and CIFAR-10 datasets.
Dataset Splits No The paper mentions training models on MNIST and CIFAR-10 and evaluating on test data, but it does not explicitly specify the proportions for training, validation, and test splits, nor does it explicitly state the use of standard splits for these datasets.
Hardware Specification No The paper does not provide specific hardware details such as GPU models, CPU types, or memory specifications used for running the experiments.
Software Dependencies No The paper describes implementation details and training strategies, but it does not specify software dependencies (e.g., libraries, frameworks) along with their version numbers.
Experiment Setup Yes For training on CIFAR-10, we choose (ฮท, ฯต, ฮปmin) = (0.05, 0.6, 0.01). The rest of the training details are as follows. We use the modified cross-entropy loss function from (Prach & Lampert, 2022), ... We use ฯ„ = 0.25. Furthermore, we train our models for 1000 epochs with a batch size of 256 with a cosine annealing strategy with an initial learning rate of 10 4 and a final learning rate 10 5, and report the average results on two seed in Table 1.