Beyond probability partitions: Calibrating neural networks with semantic aware grouping

Authors: Jia-Qi Yang, De-Chuan Zhan, Le Gan

NeurIPS 2023 | Conference PDF | Archive PDF | Plain Text | LLM Run Details

Reproducibility Variable Result LLM Response
Research Type Experimental Experimental results demonstrate that our approach achieves significant performance improvements across multiple datasets and network architectures, thus highlighting the importance of the partitioning function for calibration.
Researcher Affiliation Academia Jia-Qi Yang De-Chuan Zhan Le Gan State Key Laboratory for Novel Software Technology Nanjing University, Nanjing, 210023, China
Pseudocode Yes Algorithm 1 Train group calibration with temperature scaling (GC+TS)
Open Source Code Yes Code and Appendix are available at https://github.com/ThyrixYang/group_calibration
Open Datasets Yes To evaluate the performance of our method under various circumstances, we selected three datasets: CIFAR10, CIFAR100[31], and Imagenet[1].
Dataset Splits Yes We randomly partitioned a validation set Dval from the standard training set: CIFAR10 and CIFAR100 adopted 10% of the data for validation, while Imagenet utilized 5%.
Hardware Specification No The paper does not provide specific details about the hardware used for the experiments (e.g., GPU models, CPU types).
Software Dependencies No The paper does not provide specific version numbers for software dependencies used in the experiments.
Experiment Setup Yes The hyperparameters of the comparative methods were tuned based on the corresponding literature with 5-fold cross-validation on the CIFAR10-Resnet152 dataset. We fixed the number of groups at K = 2 and the number of partitions at U = 20, although 20 is not necessarily the optimal value. The strength of regularization was set to λ = 0.1, following a similar tuning approach as the comparative methods.