Beyond Pinball Loss: Quantile Methods for Calibrated Uncertainty Quantification

Authors: Youngseog Chung, Willie Neiswanger, Ian Char, Jeff Schneider

NeurIPS 2021 | Conference PDF | Archive PDF | Plain Text | LLM Run Details

Reproducibility Variable Result LLM Response
Research Type Experimental We provide a thorough experimental evaluation of our methods, which includes a high dimensional uncertainty quantification task in nuclear fusion.
Researcher Affiliation Academia Youngseog Chung Machine Learning Department Carnegie Mellon University Pittsburgh, PA 15213 youngsec@cs.cmu.edu Willie Neiswanger Department of Computer Science Stanford University Stanford, CA 94305 neiswanger@cs.stanford.edu Ian Char Machine Learning Department Carnegie Mellon University Pittsburgh, PA 15213 ichar@cs.cmu.edu Jeff Schneider Robotics Institute Carnegie Mellon University Pittsburgh, PA 15213 schneide@cs.cmu.edu
Pseudocode Yes Algorithm 1 MAQR 1: Input: Train data {xi, yi}N i=1, trained regression model ˆf(x) 2: Calculate residuals ϵi = yi ˆf(xi), i [N], and denote the residual dataset R = {xi, ϵi}N i=1 3: Initialize D 4: for i = 1 to N do 5: Di CONDQUANTILESESTIMATORS(R, i) (Algorithm 2) 6: D D Di 7: end for 8: Use D to fit a regression model ˆg ˆg : (x, p) 7 ϵ 9: Output: ˆf + ˆg
Open Source Code Yes Code is available at https://github.com/Youngseog Chung/calibrated-quantile-uq.
Open Datasets Yes We demonstrate the performances of our proposed methods on the standard 8 UCI datasets [2], and on a real-world problem in nuclear fusion.
Dataset Splits Yes For each dataset, we use 80% of the data for training, 10% for validation, and 10% for testing.
Hardware Specification No The paper does not specify the hardware used for running experiments, such as GPU models, CPU types, or cloud computing instances.
Software Dependencies No The paper mentions using Python-based tools and neural networks but does not provide specific version numbers for software dependencies like Python, PyTorch/TensorFlow, or CUDA.
Experiment Setup Yes Our neural network models consist of 3 hidden layers with 128 neurons each, and use ReLU activation functions for the hidden layers. We use Adam optimizer with learning rate 1e-3, batch size 64, and weight decay 1e-5.