Uploaded image for project: 'Other Specification Feedback'
  1. Other Specification Feedback
  2. OTHER-2738

Transparency versus Determinism

XMLWordPrintableJSON

    • Icon: Change Request Change Request
    • Resolution: Persuasive with Modification
    • Icon: Medium Medium
    • AI Data Lifecycle (Phase I: Data Used for Machine Learning Models) (OTHER)
    • 1.0-ballot
    • Electronic Health Record
    • Background
    • Transparency vs Determinism
    • Hide

      Thank you for your well-considered issue. The verbiage in concerned lines (130-139 and 160-167) has been edited and an updated section about Explainable AI (XAI) has been added.

      Lines 133-142 have been changed to reference the new section on Determinism/ Repeatability. Under Characteristics of Predicative AI, new lines 149-152 state: Deterministic / consistent: As defined below (see “Repeatability / Determinism”), predictive AI models, once trained, tend to be deterministic, but they may also be non-deterministic and many use non-deterministic training techniques that mean two similarly trained models may not agree.
      Lines 160-167 have also been changed to lines 174-175 to say: o Lack of determinism: As defined below, see “Repeatability / Determinism,” Generative AI tends to be non-deterministic in nature.
      New section in lines 202-359 discusses ""Explainable AI (XAI)"", replacing section on Transparency and Determinism that started on line 182. Includes sub-section on Repeatability/Determinism.

      Project team will be happy to provide most recent draft of the White Paper Informative Guide upon request and to discuss these changes; otherwise, the team requests withdrawal of Negative vote.

      Show
      Thank you for your well-considered issue. The verbiage in concerned lines (130-139 and 160-167) has been edited and an updated section about Explainable AI (XAI) has been added. Lines 133-142 have been changed to reference the new section on Determinism/ Repeatability. Under Characteristics of Predicative AI, new lines 149-152 state: Deterministic / consistent: As defined below (see “Repeatability / Determinism”), predictive AI models, once trained, tend to be deterministic, but they may also be non-deterministic and many use non-deterministic training techniques that mean two similarly trained models may not agree. Lines 160-167 have also been changed to lines 174-175 to say: o Lack of determinism: As defined below, see “Repeatability / Determinism,” Generative AI tends to be non-deterministic in nature. New section in lines 202-359 discusses ""Explainable AI (XAI)"", replacing section on Transparency and Determinism that started on line 182. Includes sub-section on Repeatability/Determinism. Project team will be happy to provide most recent draft of the White Paper Informative Guide upon request and to discuss these changes; otherwise, the team requests withdrawal of Negative vote.
    • Gary Dickinson / John Ritter: 5-0-0
    • Clarification
    • Non-substantive
    • 1.0-ballot

      I have been around AI a lot, but I have never heard the so-called common misconception that "that AI transparency is useless because AI is “non-deterministic". What is the source of this statement? It is also a very strange dichotomy. Transparency is not "about predictability". It is about having information on the data, data preprocessing, feature selection, training, testing, algorithms, etc. The comparison seems odd, and I'm not sure why the paper is putting an emphasis on output determinism in the first place.

            Unassigned Unassigned
            Mark_Kramer Mark Kramer
            Watchers:
            2 Start watching this issue

              Created:
              Updated:
              Resolved: