Parametric ρ-Norm Scaling Calibration

Authors: Siyuan Zhang, Linbo Xie

AAAI 2025 | Venue PDF | Archive PDF | Plain Text | LLM Run Details

Reproducibility Variable Result LLM Response
Research Type Experimental We evaluate our methods on multiple DNNs, including Res Net and VGG series. Our experiments are conducted on SVHN, CIFAR-10/100, 102 Flower, and Tiny-Image Net for post-hoc calibration performance. Different ablation experiments are designed to evaluate efficiency of the ρ-Norm Scaling calibration structure and the multi-level objective. In tables, the best results and relative improvements over 2nd best result in each section are in bold. Results are averaged over five runs with different seeds. Baselines: In experiments, we compare our methods with different calibration methods, such as non-parametric Hist. Binning, TS, Vector Scaling (Niculescu-Mizil and Caruana 2005).
Researcher Affiliation Academia Siyuan Zhang, Linbo Xie* School of Internet of Things Engineering, Jiangnan University EMAIL, xie EMAIL
Pseudocode Yes Algorithm 1: ρ-Norm Scaling Post-hoc Calibrator
Open Source Code No The paper does not contain any explicit statement about releasing source code, nor does it provide any links to a code repository or mention code in supplementary materials.
Open Datasets Yes Our experiments are conducted on SVHN, CIFAR-10/100, 102 Flower, and Tiny-Image Net for post-hoc calibration performance.
Dataset Splits No The paper mentions several datasets (SVHN, CIFAR-10/100, 102 Flower, Tiny-Image Net) but does not provide specific details about how these datasets were split into training, validation, or test sets, such as percentages or sample counts.
Hardware Specification No The paper does not provide specific details about the hardware used for running the experiments, such as GPU or CPU models.
Software Dependencies No The paper does not specify any software dependencies with version numbers (e.g., Python, PyTorch, TensorFlow versions).
Experiment Setup Yes In all experiments for CIFAR-10/100, the learning rate was set to 0.1, the momentum to 0.9, the weight clipping to Norm=3, and the batch size to 128. The learning rate decreased to 10% at 40% and 80% of the iterations. The weight decay was set to 10^-4 and the iteration number was 200. For the Tiny-Image Net, the learning rate was set to 0.01 and batch size was 64. The hyperparameter α is set to 1.