Information-theoretic Generalization Analysis for Expected Calibration Error

Authors: Futoshi Futami, Masahiro Fujisawa

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

Reproducibility Variable Result LLM Response
Research Type Experimental Experiments using deep learning models show that our bounds are nonvacuous thanks to this information-theoretic generalization analysis approach.
Researcher Affiliation Collaboration Futoshi Futami Osaka University / RIKEN AIP futami.futoshi.es@osaka-u.ac.jp Masahiro Fujisawa RIKEN AIP masahiro.fujisawa@riken.jp
Pseudocode No The paper does not contain any clearly labeled pseudocode or algorithm blocks. It provides theoretical analyses and mathematical derivations.
Open Source Code Yes We submitted our source codes through Open Review.
Open Datasets Yes We further conducted two binary classification tasks on MNIST [25] using a convolutional neural network (CNN) and on CIFAR-10 [21] using Res Net.
Dataset Splits No The paper describes the use of training and test datasets but does not explicitly mention a validation set split percentage or count for experiments.
Hardware Specification Yes We used NVIDIA GPUs with 32GB memory (NVIDIA DGX-1 with Tesla V100 and DGX-2) for MNIST (SGLD) and CIFAR-10 experiments. We also used CPU (Apple M1) with 16GB memory for the other experiments.
Software Dependencies No The paper mentions using 'sklearn.feature_selection.mutual_info_classif function' but does not provide specific version numbers for this or other software libraries or dependencies. It states adapting code from a previous work but does not list its dependencies with versions.
Experiment Setup Yes Optimizer Adam with 0.001 learning rate and β1 = 0.9 SGLD with 0.004 learning rate (decaying by a factor 0.9 after each 100 iterations) Batch size 128 (for Adam) or 100 (for SGLD) Num. of training samples [75, 250, 1000, 4000] Num. of epochs 200