On the Limitations of Temperature Scaling for Distributions with Overlaps

Authors: Muthu Chidambaram, Rong Ge

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

Reproducibility Variable Result LLM Response
Research Type Experimental Lastly, in Section 5 we show that our theoretical results accurately reflect practice by considering both synthetic data and image classification benchmarks. 5 EXPERIMENTS
Researcher Affiliation Academia Muthu Chidambaram Department of Computer Science Duke University muthu@cs.duke.edu Rong Ge Department of Computer Science Duke University rongge@cs.duke.edu
Pseudocode No The paper does not contain any sections or figures explicitly labeled 'Pseudocode' or 'Algorithm'.
Open Source Code Yes Code used to generate all of the plots in this paper can be found in the associated Github Repository: https://github.com/2014mchidamb/temp-scaling-limitations.
Open Datasets Yes We can also verify that the phenomena observed in synthetic data translates to the more realistic benchmarks of CIFAR-10, CIFAR-100, and SVHN.
Dataset Splits Yes For each training dataset considered in this section, we set aside 10% of the data for calibration.
Hardware Specification Yes on a single A5000 GPU
Software Dependencies No The paper mentions 'Py Torch' but does not specify a version number for it or any other software dependencies.
Experiment Setup Yes All models were trained for 200 epochs using Adam (Kingma & Ba, 2015) with the standard hyperparameters of β1 = 0.9, β2 = 0.999, a learning rate of 0.001, and a batch size of 500