Is Epistemic Uncertainty Faithfully Represented by Evidential Deep Learning Methods?

Authors: Mira Juergens, Nis Meinert, Viktor Bengs, Eyke Hüllermeier, Willem Waegeman

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

Reproducibility Variable Result LLM Response
Research Type Experimental This paper presents novel theoretical insights of evidential deep learning, highlighting the difficulties in optimizing second-order loss functions and interpreting the resulting epistemic uncertainty measures. With a systematic setup that covers a wide range of approaches for classification, regression and counts, it provides novel insights into issues of identifiability and convergence in second-order loss minimization, and the relative (rather than absolute) nature of epistemic uncertainty measures. This paper presents novel theoretical results that put previous papers in a unifying perspective by analyzing a broad class of exponential family models that cover classification, regression and count data. For the second research question, we will present in Sections 3.2–3.4 and Section 4 novel theoretical and experimental results.
Researcher Affiliation Academia 1Department of Data Analysis and Mathematical Modeling, Ghent University, Belgium 2Institue of Communications and Navigation, German Aerospace Center (DLR), Neustrelitz, Germany 3Department of Informatics, University of Munich (LMU), Germany.
Pseudocode No No pseudocode or algorithm blocks were found within the paper.
Open Source Code Yes The code for reproducing the experiments is available on Git Hub.2
Open Datasets No We generate synthetic training data DN = {(xi, yi)}N i=1 of size N, with xi Unif([0, 0.5]), and yi Bern θ(xi) . The Bernoulli parameter θ is described as a function of the one-dimensional features xi: θ(xi) = 0.5 + 0.4 sin 2πxi, thus θ(xi) [0.5, 0.9]. That is, we generate datasets {(xi, x3 i + ϵ)}N i=1 of different sample sizes, where the instances xi U([ 4, 4]) are uniformly distributed and ϵ N(0, σ2 = 9).
Dataset Splits No The paper describes generating synthetic training data and resampling it to estimate a reference distribution, but it does not specify explicit training/validation/test splits with percentages, absolute counts, or predefined partition files.
Hardware Specification No The paper does not specify the exact hardware used for experiments, such as GPU models, CPU types, or memory.
Software Dependencies No The paper mentions the use of 'Adam optimizer' but does not provide specific version numbers for any software dependencies or libraries (e.g., Python, PyTorch, TensorFlow versions).
Experiment Setup Yes We use the same model architecture for first and second-order risk minimization, consisting of 32 neurons and 2 fully-connected hidden layers. The models are trained using the Adam optimizer with a learning rate of 0.0005 and 5000 training epochs. For the regularized models, we use negative entropy of the predicted distribution with λ = 0.01.