1. Remembering and Learning

2. Recap of Chapter 3
Mostly puzzle problems requiring little knowledge to solve.
“The simplicity we discovered there was largely a simplicity of process and a simplicity of the architecture of the mind.”
Simple process
Few things to know and not many potential actions
Simple architecture
Few parameters, such as capacity of STM and storage time for LTM
What do we see when we examine complex domains

3. Semantically-Rich Domains
Simon assumes “small repertory of information processes of the sorts described in the last chapter”
Using
Input from eyes, ears, touch, etc.
Output via legs, hands, tongue
Large store of correct and incorrect information about that world
Need to understand memory

4. Long-Term Memory LTM as a library
Vast collection of information organized by topic/theme/etc.
Cross-references between items (associations)
Elaborate index enabling recognition
LTM operates as second environment that
Can be searched for information
Can be the focus of reflection/reaction

5. Outer and Inner Search Two diagnostic practices of physicians
Direct recognition – presence of symptom(s) leads to hypothesis
Search – similar to description of general problem solving (e.g. use evidence to select additional tests to gather more information)
Search is conducted simultaneously in mind of physician and body of patient
How is this like McGuckins?
Intuition as recognition, or compiled knowledge?

6. How Much Information?
A professional’s knowledge is adequate when she knows about as much as other professionals in her domain.
Time available to acquiring and maintaining knowledge will affect limit for large domains
Describing expertise
~50,000 chunks across disciplines, or 10 years of learning
When a domain exceeds this
It will increase use of externalized information stores
It will divide into subfields (specialize)
Science proceeds through producing new knowledge and compressing old through more general theories

7. Alternative Representations
Stored as data
Memory of information
Can think of databases or indexed document stores
Examples of data you know as a CS expert?
Stored as process
Memory of processes for determining information
Can think of production rules
Examples of process you know as a CS expert?

8. Understanding and Representation
Representing word problems
Did you recognize the Tea Ceremony problem?
As described, it is about the assignment of tasks to actors and constraints on how tasks can be transferred
Early programs that “understood”
UNDERSTAND – basic NLP for creating objects and relations from text
ISAAC – filling in existing schema based on parsed text
Does vast size of store necessitate complexity
Can still be thought of as simple process on big data
But increase in size can increase the number of types of relations, constraints, and interactions that must be considered during problem solving

9. Learning
Distinguish between acquiring information (data) and acquiring skills (process)
Learning by rote vs. learning with understanding
Rote learning can be regurgitated but not used as tool
Learning with understanding is faster, lasts longer, and is more generally applicable
Differences in indexing, redundancy, and representation
How many stars were on the US flag in 1940?
How did you answer that?

10. Production (rule-based) Systems
Easier to generate new data representations and add data than to add new processes
Programming is hard
Production systems use simple encoding of process as data
Condition -> Action
Production rules can be used in multiple ways
Governed by perception (stimulus or data driven)
Forward chaining
Governed by goal (goal driven)
Often backward chaining
Adaptive production system – where productions are added, deleted, or modified during runtime
Writing code that writes code
Writing code that modifies itself?

11. Learning from Examples
Can take steps in example and generalize to process description (e.g. production rules)
How to generalize at right level?
Learning by doing for learning process
Discussions of understanding and learning relate to constructivist educational pedagogy
Learning with understanding better than learning by rote
+ Learning process more applicable than learning data
= Learning by example/doing

12. Discovery Discovery is a type of learning
“what constitutes novelty depends on what is already in the mind of the problem solver”
Discovery does not have a goal
AM used criteria for judging how interesting a pattern is to produce findings
BACON located monotonically varying data and introduced new concepts to represent relations found
DENDRAL and MECHEM generated publishable results (was this the beginning of “Big Data”?)
Success in discovery is related to selected representation
Focus of attention determined by representation
E.g. mutilated checker board

13. Simon Sticks to His Guns
He continues to argue that human beings can be viewed as simple problem solving systems
Provided that
“we include in what we call the human environment the cocoon of information, stored in books and in long-term memory, that we spin about ourselves.”