Calibrated and Sharp Uncertainties in Deep Learning via Density Estimation

Authors: Volodymyr Kuleshov, Shachi Deshpande

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

Reproducibility Variable Result LLM Response
Research Type Experimental Our results yield empirical performance improvements on linear and deep Bayesian models and suggest that calibration should be increasingly leveraged across machine learning. and Empirically, we find that our method consistently outputs well-calibrated predictions in linear and deep Bayesian models, and improves performance on downstream tasks with minimal implementation overhead.
Researcher Affiliation Academia 1Department of Computer Science, Cornell Tech and Cornell University, New York, NY.
Pseudocode Yes Algorithm 1 Distribution Recalibration Framework and Algorithm 2 Distribution Calibrated Regression and Algorithm 3 Distribution Calibrated Classification
Open Source Code No The paper does not include an explicit statement about open-sourcing code or a link to a code repository.
Open Datasets Yes Datasets. We use a number of UCI regression datasets varying in size from 194 to 8192 training instances; each training input may have between 6 and 159 continuous features. ... We also perform classification on the following standard datasets: MNIST, SVHN, CIFAR10.
Dataset Splits Yes We randomly use 25% of each dataset for testing, and use the rest for training. We held out 15% of the training set (up to max of 500 datapoints) for recalibration.
Hardware Specification No The paper does not provide any specific hardware details such as GPU/CPU models, memory, or cloud instances used for running the experiments.
Software Dependencies No The paper mentions 'implemented easily within deep learning frameworks' but does not specify any software names with version numbers (e.g., Python, PyTorch, TensorFlow, CUDA).
Experiment Setup Yes In our UCI experiments, we use fully-connected feedforward neural networks with two layers of 128 hidden units with a dropout rate of 0.5 and parametric Re LU non-linearities. ... Our recalibrator R was also a densely connected neural network with two fully connected hidden layers of 20 units each and parametric Re LU non-linearities. ... In regression experiments, we featurized input distributions F using nine quantiles [0.1, ..., 0.9]. We trained R using the quantile regression objective of Algorithm 2;...