What is a Knowledge Representation?

(and related matters)

Bijan Parsia

Knowledge and Representation

• Knowledge
  • True justified belief (plus a little...)
  • We are more doing belief representation (see role 1)
• Representation
  • A depiction...(whoops)
  • A encoding...(damn)
  • A symbolic structure which bears correspondence relationships with states of affairs which are interpreted (argh) or manipluated in ways that are sensitive to those relationships
  • Ok, that hurt
• So, What is a KR?

Five Roles

1. Surrogate
   • That is, a representation
2. Expression of ontological commitment
   • of the world
3. Theory of intelligent reasoning
   • and our knowledge of it
4. Medium of efficient computation
   • that is accessible to programs
5. Medium of human expression
   • and usable

A Surrogate

• Is a representation a surrogate?
• Simulation/picture theory of meaning
• Representations are abstractions
• No representation captures everything
• It wouldn't be useful it if did!
• The goodness of a representation is context sensitive and interest relative
• It depends on the applications
• We might trade off
• Accuracy for intelligibiliy
• Detail for performance
• What does the representation actually represent?
  

Consider the foci! (1)

• (Commonsense, scientific, conceptual modeling)
• Different concerns lead to different representations
• Representing objects in a room for a robot
• Representing objects for an invetory system
• Representing proteins for a gene classifier
• Representing protiens for a nutrition program

Ontological Commitment

• Imperfection of reps entails choice of what to represent
• And how to
• Even if we retained everything, organization and emphasis matter!
• 5th fallacy of metadata
• Representation is relational
• See surrogacy
• Formalisms (typically) constrain, not determine the relations
• Model theory
• The grounding problem
• Don't be seduced by "ontology"
• This isn't philosophy
• Ontological talk can be convenient

Consider the foci! (2)

• Each focus has different families of committment
• Commonsense is concerned with "middle sized dry goods"
• Specialized areas have more complex and fine grained ontologies 
• But often less breath

"A KR is not a data structure"

• Critical point (analogy with DBs)
  • ER diagrams (P.P. Chen) -- 4 levels of views of data:
    1. "Information concerning entities and relationships which exist in our minds."
       
    2. "Information structure -- organization of information in which entities and relationships are represented by data. "
    3. "Access-path-independent data structure -- the data structures which are not involved with search schemes, indexing schemes, etc."
    4. "Access-path-dependent data structure."
• Data structures implement the representations
  • And there is a lot of choice
  • Many details of the structures play no representational role
• Choices at every level affect the success of an application

Intelligent Reasoning

• Representation and reasoning are correlative
• Tendentious, but reasonable
• Representation without manipulation (or manipulatabilty) isn't a representation
• If you can't use it to acquire information, in what sense does it represent?
• Components
• Fundamental conception
• Deduction, human behavior, etc. etc.
• Sanctioned inferences
• Recommended inferences
• We primarily focus on deduction or, more loosely, entailment

Consider the foci! (3)

• Commonsense reasoning 
• Lot of jumping to conclusions
• Fast reasoning
• Good enough reasoning
• Scientific reasoning
• Highly mathematical (sometimes)
• Getting it right more important

Strong Cognitive Adequacy

• Should the conception be a model of human reasoning?
• Usually associated with the psychological or neuroscience or commonsense reasoning traditions
• KRs can help us understand human intelligence
• Flipside, since we are the most successful cognitive agents we know, aping us might be successful
• Should KRs be strongly cognitive adequate?
• Counterpoints
• Human intelligence isn't the only form of intelligence
• And it's not reliable
• Information managment isn't (typically) cognition
• Think libraries!
• Aping us might not be sound engineering
• Drive any bipedal cars today?

Efficient Computation

• Programs have to work with representations
• The representation management system is a component in a larger system
• If the repman system is inefficient, programmers  will compensate
• (There's a Strong Cognitive Adquacy to Efficient Computation fallacy)
• Most interesting systems work at "large" scale
• Even without the Cyc hypothesis
• Consider databases
• Representations get complex quickly
• People need prosthetics to work well with them

Human Expression

• Humans interact with representations
• Effectively
• At least of certain kinds
• People must work with KRs
• Generating them
• Using them to build systems
• Using them when using the system
• Mycin explanations
• Weak Cognitive Adquacy
• KRs must be usable

Consider the foci! (5)

• Scientific KR must make sense to practitioners
• Conceptual modeling must make sense to programmers
• If the system is opaque, then experts cannot help
• If the system is (too) opaque, then users may not be inclined to trust the system
• Consider medical diagnosis
• Or loan approval
• Commonsense varies
• If we're just building something "intelligent" e.g., a robot that can navigate a room, we might not care too much about the usability of the representation