Localized Adaptive Risk Control

Authors: Matteo Zecchin, Osvaldo Simeone

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

Reproducibility Variable Result LLM Response
Research Type Experimental In Section 3 we showcase the superior conditional risk control properties of L-ARC as compared to ARC for the task of electricity demand forecasting, tumor segmentation, and beam selection in wireless networks.
Researcher Affiliation Academia Matteo Zecchin Osvaldo Simeone Centre for Intelligent Information Processing Systems Department of Engineering King s College London London, United Kingdom {matteo.1.zecchin,osvaldo.simeone}@kcl.ac.uk
Pseudocode No The paper does not contain structured pseudocode or algorithm blocks. The update rules are described as formulas within the text.
Open Source Code Yes The simulation code is available at https://github.com/kclip/localized-adaptive-risk-control.git.
Open Datasets Yes Firstly, we address the task of electricity demand forecasting, utilizing data from the Elec2 dataset [Harries et al., 1999]. Next, we present an experiment focusing on tumor segmentation, where the data comprises i.i.d. samples drawn from various image datasets [Jha et al., 2020, Bernal et al., 2015, 2012, Silva et al., 2014, Vázquez et al., 2017]. We reserve 50 samples from each repository for testing the performance post-calibration, while the remaining T = 2098 samples are used for online calibration. In this section, we consider an image classification task under calibration requirements based on the fruit-360 dataset [Muresan and Oltean, 2018].
Dataset Splits No The paper mentions training and testing data but does not explicitly describe validation splits or cross-validation setups. The term "validation" is used in the context of the prediction sets, not data splitting.
Hardware Specification Yes All the experiments are conducted on a consumer-grade Mac Mini with an M1 chip.
Software Dependencies No The paper mentions specific models like Res Net and Pra Net but does not provide specific version numbers for software libraries or dependencies used for implementation or experimentation.
Experiment Setup Yes Unless stated otherwise, we instantiate L-ARC with the RBF kernel k(x, x ) = κ exp( x x 2 /l) with κ = 1, length scale l = 1 and regularization parameter λ = 10 4. Both ARC and L-ARC use the learning rate ηt = t 1/2. L-ARC is instantiated with the RBF kernel k(x, x ) = κ exp( ϕ(x) ϕ(x ) 2 /l), where ϕ(x) is a 7-dimensional feature vector corresponding to the daily average electricity demand during the past 7 days. Both ARC and L-ARC are run using the same decaying learning rate ηt = 0.1t 1/2. L-ARC is instantiated with the RBF kernel k(x, x ) = κ exp( ϕ(x) ϕ(x ) 2 /l), where ϕ(x) is a 5-dimensional feature vector obtained via the principal component analysis (PCA) from the last hidden layer of the Res Net model used in Pra Net.