Uncertainty, complexity and model confidence
The bibliography cites a theory of AI uncertainty built on Rademacher complexity and Shannon entropy, and the dropout-as-Bayesian-approximation method for representing model uncertainty in deep learning.
The requirement in plain English
The bibliography cites a theory of AI uncertainty built on Rademacher complexity and Shannon entropy, and the dropout-as-Bayesian-approximation method for representing model uncertainty in deep learning.
What the public record says
The bibliography entries name the works. The public view gives no summaries and no clause references.
Why it matters to you
A robustness estimate without an uncertainty statement is a number without a warranty. The two citations approach uncertainty from different directions: one theoretical, through complexity and entropy, one practical, through approximate Bayesian inference at prediction time. For a buyer, the useful question is whether the method you chose reports uncertainty in a form a risk owner can use, and whether that uncertainty was estimated on the same distribution as the robustness result.
Evidence a reviewer asks for
- The uncertainty method and its assumptions.
- The uncertainty estimate reported alongside the robustness result.
- The relationship between the uncertainty sample and the robustness sample.
- A note on calibration, if the method depends on it.
Checks to run this month
- Is uncertainty reported with the result?
- Are the assumptions of the uncertainty method stated?
- Is the uncertainty calibrated on the deployment distribution?
Dates
Papers published 2016 and 2020. FDIS registered 2026-08-21.
Cross-references
- A statistical approach to robustness
- Set the acceptance threshold and the decision rule
- Assemble the evidence pack a reviewer will test
What this page does not state
The public view gives no summaries and no statement of how the draft uses the papers.