We gratefully acknowledge support from
the Simons Foundation and member institutions.
Full-text links:

Download:

Current browse context:

cs.LG

Change to browse by:

References & Citations

DBLP - CS Bibliography

Bookmark

(what is this?)
CiteULike logo BibSonomy logo Mendeley logo del.icio.us logo Digg logo Reddit logo ScienceWISE logo

Computer Science > Machine Learning

Title: Looking at the posterior: on the origin of uncertainty in neural-network classification

Abstract: Bayesian inference can quantify uncertainty in the predictions of neural networks using posterior distributions for model parameters and network output. By looking at these posterior distributions, one can separate the origin of uncertainty into aleatoric and epistemic. We use the joint distribution of predictive uncertainty and epistemic uncertainty to quantify how this interpretation of uncertainty depends upon model architecture, dataset complexity, and data distributional shifts in image classification tasks. We conclude that the origin of uncertainty is subjective to each neural network and that the quantification of the induced uncertainty from data distributional shifts depends on the complexity of the underlying dataset. Furthermore, we show that the joint distribution of predictive and epistemic uncertainty can be used to identify data domains where the model is most accurate. To arrive at these results, we use two common posterior approximation methods, Monte-Carlo dropout and deep ensembles, for fully-connected, convolutional and attention-based neural networks.
Comments: 25 pages, 6 figures, 5 tables, 1 appendix
Subjects: Machine Learning (cs.LG); Computer Vision and Pattern Recognition (cs.CV); Machine Learning (stat.ML)
Cite as: arXiv:2211.14605 [cs.LG]
  (or arXiv:2211.14605v1 [cs.LG] for this version)

Submission history

From: Bernhard Mehlig [view email]
[v1] Sat, 26 Nov 2022 16:13:32 GMT (861kb,D)

Link back to: arXiv, form interface, contact.