Efficient and Differentiable Conformal Prediction with General Function Classes

Authors: Yu Bai, Song Mei, Huan Wang, Yingbo Zhou, Caiming Xiong

ICLR 2022 | Conference PDF | Archive PDF | Plain Text | LLM Run Details

Reproducibility Variable Result LLM Response
Research Type Experimental We empirically test our CP-Gen-Recal algorithm (using the practical implementation (8)) on three representative real-data tasks.
Researcher Affiliation Collaboration Yu Bai Salesforce Research yu.bai@salesforce.com Song Mei UC Berkeley songmei@berkeley.edu Huan Wang & Yingbo Zhou & Caiming Xiong Salesforce Research {huan.wang, yingbo.zhou, cxiong}@salesforce.com
Pseudocode Yes Algorithm 1 Conformal Prediction with General Function Class (CP-Gen) and Algorithm 2 Conformal Prediction with General Fn. Class and Recalibration (CP-Gen-Recal).
Open Source Code Yes Code available at https://github.com/allenbai01/cp-gen.
Open Datasets Yes Our choice of the datasets follows (Feldman et al., 2021). We provide information about these datasets in Table 3. ... Physicochemical properties of protein tertiary structure data set. https://archive.ics.uci.edu/ml/datasets/Physicochemical+Properties+of+Tertiary+Structure. Accessed: January, 2019.
Dataset Splits Yes All datasets are standardized so that inputs and labels have mean 0 and standard deviation 1, and split into (train, cal, recal, test) with size 70%, 10%, 10%, 10% (varying with the random seed).
Hardware Specification No The paper does not provide specific details about the hardware (e.g., GPU models, CPU types) used for running the experiments.
Software Dependencies No The paper mentions software components like '3-layer MLP', 'momentum SGD', and 'Open AI Gym' but does not specify their version numbers or other software dependencies with specific versions.
Experiment Setup Yes Our network architecture is a 3-layer MLP with width 64 and output dimension 2 (for the lower and upper quantile). We use momentum SGD with initial learning rate 10 3 and momentum 0.9, batch-size 1024, and run the optimization for a max of 10000 epochs. A 10x learning rate decay is performed if the validation loss on Dcal has not decreased in 10 epochs, and we stop the learning whenever the learning rate decay happens for 3 times. ... To solve that optimization problem, we use SGD on (θ, t) with learning rate 0.01 and (ascent on) λ with learning rate 0.1. The batch-size here is 256 and the number of episodes is 1000.