Main Page
AI KR (Artificial Intelligence Knowledge Representation) Community Group Wiki
These pages are used by members of the AI KR (Artificial Intelligence Knowledge Representation) Community Group for collaborative work.
Please note that the original AI KR CG landing page cannot be edited and may be slighthly out of date as the mission may have shifted since the group started (feature request?)
In brief: Knowledge Representation is a broad topic that can be tackled from a variety of perspectives and approaches. Classical AI KR textbooks (Brachman, Levesque and many others) have been laying the foundations for KR in computer science and AI since the 50's, and although the core principles and underlying concepts in the domain have remained the same, the field of AI has been evolving fast and has since become central to the world wide web and intelligent information systems.
In recent decades (times goes fast) the overall notion of KR has been somewhat cast aside, or rather, it has transformed into Conceptual Modelling and Ontology Engineering. Many members of this CG have been fortunate over the years to exchange meaningfully on related topics, and to observe this evolution first hand.
This CG was started (2018) following a surge of renewed interest in AI, with notable lack of attention paid to shared and explicit KR, especially in ML (Machine Learning). Through posts on the group's mailing list, and other contributions, members of this CG try to follow closely the developments in AI identifying the key role that KR plays in addressing open issues, especially systemic risks which can become embedded in process chains (we should talk more about these actually).
The main resource in use is the AI KR Mailing List, an open public archive, where key resources, ideas and discussions are shared informally. Please view the archive by thread to get a sense of what we are talking about, what helps us keep our heads above water, feel free to join and pitch in. If having issues joining, please contact the chairs.
Contributions to date Between 2015 and 2018 ML (Machine Learning) started being used to drive advances in automation systems designed as rather unpredictable and unreliable black boxes, without explicit knowledge representation (KR) which was dismissed as irrelevant to ML (considered connectionist AI and a different thing from symbolic KR).
The lack of replicable results in ML is ascribed to stochasticism. probabilic and statistical approaches, but they can be rather unpredictable, underliable and unexplainable. That's when AI can really get out of hand.
Soon enough XAI (explainable AI) became a thing, and this list contributed to the understanding that actually, it is only through explicit and shared KR that AI can be explainable.
The link between KR and XAI has been made in literature as well as in presentations and publications and shared as research notes, such as NIST. to be updated
We are confident that the discussions and resources shared here contribute to bringing AI back into the hands of humans, and KR back into the scholarly and research agendas in AI. The relevance to ML is now understood and widely accepted.
At a minimum, people have started paying attention and doing what we say, although regrettably often without reference nor proper attribution.
Historical gaps in teaching and learning KR have been reiterated and brought to light, and the case for the importance of having explicit and shared KR in symbolic AND subsymbolic AI is now generally accepted but not yet formally inlcuded in teaching curricula.
The CG participated in W3C TPAC 2021 (thank you W3C) produced a report summarizing relevant activities and discussions.
The Concept Maps seem in good order (Thanks Viorel and CJ) also Zotero KR Library Thank you Mike B
Long Term Goals The plan is to consolidate the contribution with further resources and possibly a web standard for Neuro-symbolic Integration. The group is also open to expand its long term goals and direction So feel free to pitch in.
Please join our HuggingFace community to work on model cards if you manage to work out how to join (IQ test>)
Wishlist for 2024: 1. Decide on primary goals of the CG (say for example, working towards a web standard for neurosymbolic integration? other?)
2. Make sure the resources relevant to support the stated goals are open access (if not yet available, work towards making open access version of materials behind paywall)
3. Update and link prominently on the main wiki page the resources which are good enough
4. Update and organise usable resources, check that links are alive and related resources and archive all the half dead not maintained ones
Acknowledgements and Apologies
The CG thanks everyone who is contributing to the resource and putting effort in creating and maintaining the discussions alive and these wiki pages Apologies for limited coherence and out of date resources and pages at times and for linking to resources beyond a paywall. We aim to produce open access versions of all resources linked here. It is work in progress.
Resources
THE SECTIONS BELOW NEED REVISION AND UPDATING - PLEASE IGNORE OR TAKE WITH A PING OF SALT. BEAR WITH US WHILE WE BRING OUR ACT TOGETHER
Simply put, Knowledge representation is defined by the method used to explain or decode knowledge
Meta : explains the types of Knowledge and logical reasoning used e.g. Spatial-Temporal reasoning. Heuristic : explains the knowledge of a subject expert e.g. Frame facet. Procedural : explains how to achieve a desired outcome e.g. Event-condition-action rule. Declarative : explains facts, concepts, principles and relationship to planned behavior e.g. Mathematical equation. Structural : explains types of relationship that exist between concepts/ objects e.g. Semantic Network.
The links below provide information about specific methods
Report Draft
AI KR Guideline: KR for eGovernance and Trustworthy AI
The intent of the guideline for eGovernance of Knowledge Representation for a Trustworthy AI (Artificial Intelligence) system is to promote good governance. As in, the implementation of AIKR (Artificial Intelligence Knowledge Representation) eGovernance processes and functions will enable accountable and transparent governance. It is based on:
Artificial Intelligence describes a system that mimics "cognitive" functions which humans associate with the human mind, such as "learning" and "problem solving". AIKR can be (1) meta-representation information about the AI system’s own state, and/or (2) information about a domain, in a form that enables an AI system to provide a solution for a specific problematic task.
A Trustworthy AI System is said to have three characteristics that are part of its entire life cycle: 1. Lawful: AI System complies with all applicable laws and regulations, such as, provision audit data defined by a governance operating model; 2. Ethical: AI System ensures adherence to pivotal principles, such as, confidentiality, autonomy, accountability and veracity; 3. Robust: AI System is designed to handle uncertainty and tolerate perturbation from a likely threat perspective, such as, design considerations incorporate human, social and technology risk factors
In Progress
- Shared KR for AI Commons - The heart of AI Commons is Shared KR
Paper Draft status: needs advancing
- Deep KR for EGovernance, Talk, Taipei September 2019/
- KR Meets Brain Science (Conference forthcoming)
Paper Draft, Not Shared - HAVE SLIDES! WORKING ON THE PAPER!
Feedback to other organisations
Online Meetings dates and agendas
Please link here the dates and agendas and minutes for the meetings? So that participants can also add to the agendas and edit the minutes
11 February 2020 [[1]]
Other Calls
Leveraging the StratML specification for AIKR using KAIROS Stakeholder(s): AIKR CG https://www.w3.org/community/aikr/welcome
Role: AI KR Srategist Vision AIKR Strategist uses StratML vocabulary and a KAIROS approach to monitor performance of AIKR objects implemented by machine learning powered services. Mission To determine if use of performance tracking of AIKR objects implemented by machine learning can provide the foundation for a Trustworthy AI System. Ontological Statement: life cycle characteristics of Trustworthy AI System 1. Lawful: AI System complies with all applicable laws and regulations, such as, provision audit data defined by a governance operating model; 2. Ethical: AI System ensures adherence to pivotal principles, such as, confidentiality, autonomy, accountability and veracity; 3. Robust: AI System is designed to handle uncertainty and tolerate perturbation from a likely threat perspective, such as, design considerations incorporate human, social and technology risk factors.
Goal Machine Learning Evaluation Evaluate machine learning models Stakeholder(s): https://www.w3.org/community/aikr Role: Community of Interest Objective(s): Towards building a foundation for a Trustworthy AIKR - primary objectives are (a) document the vision, values, goals, objectives for one or more AIKR objects, (b) Track AIKR object performance outcome via KPI (Key Performance Indicator) based on supervised learning models measurements; https://www.stratnavapp.com/StratML/Part2/861566c8-e9be-4642-b52f-f673fa499f4e/Styled secondary objective is to employ ontological statements when explaining AIKR object audit data, veracity facts and (human, social and technology) risk mitigation factors..
Co-leadership
All members can lead and shape the work being done by providing input in the stakeholder survey (linked to in the group's home page).