Symbolic Learning
1. The Problem
create symbolic structures to summarize, generalize, and explain observations
non-deductive inference: induction, abduction, analogy, and concept formation
learning about beliefs
2. Proposed Solutions
decision tree: a tool for classification, as for 20-question game; learning by recursive binary partition
concept learning: from instances to concepts, supervised vs. unsupervised
hypothesis generating and testing: Current-Best Learning and Version Space Learning
inductive logic programming: from concrete cases to general rules
induction and abduction as reversed deduction
abduction as explanation
the formalization of analogy
3. Issues
exception tolerance, accuracy/simplicity tradeoff, overfitting
huge hypothesis space, lack of generating procedure
4. Reading
Sections 18.1-3, 19.1-2, 19.5
5. Ideas
Learning is central to intelligence. Computer can learn.
reasonable:
- New symbolic structure can be learned.
- Learning is different from logical deduction, though related to it.
- There are multiple ways to learn.
problematic:
- the performance/learning distinction
- the generating/testing separation for hypothesis
- generalization/explanation/analogy allow no exception
- learning as problem solving, following algorithm that turns input to output
in NARS: