Computational Discovery of Communicable Knowledge

1
Architectures for Adaptive Interpretation
Pat Langley Computational Learning
Laboratory Center for the Study of Language and
Information Stanford University, Stanford,
California http//cll.stanford.edu/
2
Characteristics of Cognitive Architectures

• As typically defined and utilized, a cognitive
  architecture
• aims to demonstrate generality and flexibility,
  rather than success on a single application
  domain
• specifies the infrastructure that holds constant
  over domains
• focuses on functional structures and processes,
  rather than on the knowledge or implementation
  levels
• commits to representations and organizations of
  knowledge
• comes with a programming language for encoding
  knowledge and constructing intelligent systems.

These design principles apply not only to
architectures focused on action, but also to ones
focused on explanation and understanding.
3
Architectures for Action and Interpretation

• Most previous architectures, like Soar, ACT-R,
  Prodigy, and 3T were designed for action. As a
  result
• Short-term beliefs focus on goals and plans
• Long-term knowledge focuses on skills and
  procedures
• Inference is implemented with one-way
  production rules.
• To support flexible understanding and
  explanation, we need an alternative class of
  architectures in which
• Beliefs focus on inferences made from knowledge
  and facts
• An episodic belief memory replaces short-term
  memory
• Knowledge exists to generate accurate and
  useful inferences
• Inference involves flexible abduction rather
  than deduction.

We need architectures of this sort for robust
learning by reading.
4
Interpretive Architecture Structures and
Processes
Fixed Generator
Answers and Summaries
Learning and Revision
New General Knowledge
Old Specific Beliefs
Old General Knowledge
New Specific Beliefs
Inference and Interpretation
Fixed Parser
Statements and Questions
5
Basic Interpretive Cycle

• For each input I,
• 1. Use the fixed parser to generate new beliefs.
  
• 2. Update the old beliefs by incorporating the
  new beliefs.
• 3. Use the interpreter to infer additional new
  beliefs.
• 4. If not yet done, go to step 3 else go to
  step 5.
• 5. Use the learner to add or revise knowledge
  structures.
• 6. If not yet done, do to step 5 else go to
  step 7.
• 7. Use the fixed generator to produce output.

Inference is driven by new facts or questions,
but guided by old beliefs and knowledge learning
is driven by new beliefs but biased by old
knowledge and beliefs. This leaves open key
questions about the eagerness/laziness of
inference, the size of input/output units, and
the degree to which learning and inference are
interleaved.