Sampling-based Multi-dimensional Recalibration

Authors: Youngseog Chung, Ian Char, Jeff Schneider

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

Reproducibility Variable Result LLM Response
Research Type Experimental We demonstrate the performance of our method and the quality of the recalibrated samples on a suite of benchmark datasets in multidimensional regression, a real-world dataset in modeling plasma dynamics during nuclear fusion reactions, and on a decision-making application in forecasting demand.
Researcher Affiliation Academia 1Machine Learning Department; 2Robotics Institute, Carnegie Mellon University, Pittsburgh, PA 15213.
Pseudocode Yes Algorithm 1 HDR Recalibration: Training
Open Source Code Yes Code is available at: https://github.com/YoungseogChung/multi-dimensional-recalibration
Open Datasets Yes Datasets. The mulan benchmark (Tsoumakas et al., 2011) is a set of prediction tasks with multi-dimensional targets of up to 16 dimensions.
Dataset Splits Yes On each dataset, we make train-validation-test splits of proportions [65%, 20%, 15%]
Hardware Specification Yes All of the model training was done with 4 NVIDIA Ge Force RTX 2080 Ti GPUs. All of the evaluation was done on a CPU machine with Intel(R) Xeon(R) Gold 6238 CPU @ 2.10GHz.
Software Dependencies No The paper mentions software like 'Uncertainty Toolbox' and 'NGBoost' but does not specify version numbers for these or other key software dependencies required for replication.
Experiment Setup Yes For all of the datasets, the PNN trained has 5 fully connected layers, each with 200 hidden units, and the output parametrizes a diagonal Gaussian with a mean and a log-variance prediction. The Gaussian likelihood loss was used for training, with a learning rate of 0.001 and no weight decay was used. Training was halted early if the validation loss did not improve for more than 100 epochs, for a maximum of 1000 epochs.