From OntologPSMW

Revision as of 09:30, 24 April 2019 by Garybc (Talk | contribs)

Jump to: navigation, search
[ ]
Session Synthesis Session 1
Duration 1 hour
Date/Time 27 February 2019 17:00 GMT
9:00am PST/12:00pm EST
5:00pm GMT/6:00pm CET
Convener Ken Baclawski


Ontology Summit 2019 Synthesis Session 1     (2)

Abstract     (2A)

The aim of this week's session is to synthesize the lessons learned so far on the tracks that are under way. Each track has met once, and so we will have gained insights from a combination of invited speakers, chat log comments and blog page discussions.     (2A1)

A second synthesis session will take place after the tracks have met again, and that, along with today's session outcome, will form the basis of this year's Ontology Summit Communiqué.     (2A2)

Agenda     (2B)

Summary of Ontology Summit 2019 Sessions     (2C)

There were 9 sessions so far. Each session had the proceedings (from the chat room) and a recording (one audio recording and the rest video recordings). The following are the speakers with links to their presentation slides (when they were provided) and the recordings.     (2C1)

11/14John SowaExplanations and help facilities designed for peopleSlidesVideo
11/28Ram D. Sriram and Ravi SharmaIntroductory Remarks on XAISlidesVideo
Derek DoranOkay but Really... What is Explainable AI? Notions and Conceptualizations of the FieldSlides
12/05Gary Berg-Cross and Torsten HahmannIntroduction to Commonsense Knowledge and ReasoningSlidesVideo
1/16Ken BaclawskiIntroductory RemarksSlidesVideo
Gary Berg-Cross and Torsten HahmannCommonsenseSlides
Donna Fritzsche and Mark UnderwoodNarrativeSlides
Mark Underwood and Mike BennettFinancial Explanation
Ram D. Sriram and David WhittenMedical Explanation
Ram D. Sriram and Ravi SharmaExplainable AISlides
1/23Michael GrüningerOntologies for the Physical Turing TestSlidesVideo
Benjamin GrosofAn Overview of Explanation: Concepts, Uses, and IssuesSlides
1/30Donna FritzscheIntroduction to NarrativeAudio only
Ken BaclawskiProof as Explanation and NarrativeSlides
Mark UnderwoodBag of Verses: Frameworks for Narration from Cognitive PsychologySlides
2/6Mike BennettFinancial Explanations IntroductionSlidesVideo
Mark UnderwoodExplanation Use Cases from Regulatory and Service Quality Drivers in Retail Credit Card FinanceSlides
Mike BennettFinancial Industry ExplanationsSlides
2/13David WhittenIntroduction to Medical Explanation SystemsSlides
Augie TuranoReview and Recommendations from past Experience with Medical Explanation SystemsSlides
Ram D. SriramXAI for BiomedicineSlides
2/20William ClanceyExplainable AI Past, Present, and Future–A Scientific Modeling ApproachSlidesVideo

Conference Call Information     (2D)

Attendees     (2E)

Proceedings     (2F)

Resources     (2G)

Here are some ideas for a working synthesis outline for Explanations (Gary Berg-Cross)     (2G1)

1. Meaning of Explanation [An explanation is the answer to the question "Why?" as well the answers to followup questions such as "Where do I go from here?"] – there are range of these     (2G2)

  • Grosof deductive Proof , with a formal knowledge representation (KR) – is the gold standard, but there are many types with different representations     (2G3)
    • – E.g., natural deduction –HS geometry there is also probabilistic     (2G3A)
  • Causal model Explanations     (2G4)

There are a range of concepts related to explanation     (2G5)

Additional Aspects/Modifiers of explanation:     (2G10)

Trending-Up concepts of explanation     (2G13)

  • Influentiality – , heavily weighted hidden nodes and edges effect     (2G14)
  • Reconstruction – simpler / easier-to-comprehend model     (2G15)
  • Lateral relevance – interactivity for exploration     (2G16)
  • Affordance of Conversational human-computer interaction (HCI)     (2G17)
  • Good explanations quickly get into issue of understanding & meaning since much meaning involves background knowledge and commonsense and lies in the implicit and unspoken.     (2G18)
    • What does it mean to understand , follow and explain a set of instructions?     (2G18A)

2. Problems and issues     (2G19)

  • From GOFAI     (2G20)
    • An early goal of AI was to teach/program computers with enough factual (often commonsense) knowledge about the world so that they could reason about it in the way people do which is not strictly logical is some circumstances     (2G20A)
  • early AI demonstrated that the nature and scale of the problem was difficult.     (2G21)
    • one reason is that simple, direct approaches like rule based systems were brittle.     (2G21A)
  • People seemed to need a vast store of everyday, background knowledge for common tasks. A variety of background knowledge was needed to understand & explain decisions     (2G22)
  • Do we have a small, common ontology that we mostly all share for representing and reasoning about the physical world?     (2G23)
  1. it remains challenging to design and evaluate a software system that represents commonsense knowledge and that can support reasoning (such as deduction and explanation) in everyday tasks. (evidence from modified Physical Turing Tests)     (2G25)

3. From XAI     (2G27)

  • Performance vs. Explainability: DARPA XAI Program     (2G28)
    • While DL performs well is is very low on explainability which decision trees are the reverse     (2G28A)
    • One context of recent work is that of Deep machine-learning systems. Explanations for their decisions can be problematic since one cay say what they learn is a dimension space of numbers that have no words of any kind, so explanation in natural language is not immediately available.     (2G28B)
    • One approach to handling this problem is called Deep Explanation which uses modified deep learning techniques to learn explainable features as part of what it learns.     (2G28C)
    • More traditional GOFAI approaches may develop Interpretable Models using techniques that learn more structured, interpretable, causal models.     (2G28D)
  • Concept of Humagic Knowledge     (2G29)
  1. Bridging from sub-symbolic to symbolic -ontologies help constrain options     (2G31)

4. Application areas     (2G32)

  1. Examples of successes? Rulelog’s Core includes Restraint bounded rationality     (2G35)

5. Relevance and relation to context     (2G36)

6. Synergies with commonsense reasoning     (2G38)

7. Success stories/systems     (2G40)

  1. Issues Today in the Field of Explanation /Questions     (2G42)
  • How do we evaluate these ontologies supporting explanations and commonsense understanding?     (2G43)
  • How are these explanations ontologies related to existing upper ontologies?     (2G44)

8. Conclusions     (2G45)

computers, AIs, and robots. And that makes meanings matter even more. But it remains a hard problem.     (2G47)

sentience—the capacity to feel, perceive, or experience subjectively.     (2G49)

  • Benefits of Explanation (Grosof)     (2G50)
    • Semi-automatic decision support     (2G50A)
    • Might lead to fully-automatic decision making – E.g., in deep deduction about policies and legal – especially the business and medicine topics.     (2G50B)
    • Useful for Education and training, i.e., e-learning – E.g., Digital Socrates concept by Janine Bloomfield of Coherent Knowledge     (2G50C)
    • Accountability • Knowledge debugging in KB development     (2G50D)
    • Trust in systems – Competence and correctness – Ethicality, fairness, and legality     (2G50E)
    • Supports Human-machine interaction and User engagement (see Sowa also)     (2G50F)
    • Supports Reuse / and guide choice for transfer of knowledge     (2G50G)

9. Contemporary Issues     (2G51)

  • Confusion about concepts – Esp. among non-research industry and media – But needs to be addressed first in the research community     (2G52)
  • Mission creep, i.e., expansivity of task/aspect – Esp. among researchers. E.g., IJCAI-18 workshop on explainable AI.     (2G53)
  • Ignorance of what’s already practical – E.g., in deep policy/legal deduction for decisions: full explanation of extended logic programs, with NL generation and interactive drill-down navigation – E.g., in cognitive search: provenance and focus and lateral relevance, in extended knowledge graphs     (2G54)
  • Disconnect between users and investors     (2G55)
  • (Ignorance of past relevant work)     (2G56)
  • Some envision a fruitful marriage between classic logical approaches (ontologies) with statistical approaches which may lead to context-adaptive systems (stochastic ontologies) that might work similar to the human brain.     (2G57)

Previous Meetings     (2H)

Next Meetings     (2I)