Theories of Learning

PSY 565

Study tips in addition to the learning objectives for Symbolic Models of Knowledge

LTM encodes knowledge by meaning. The codes could be viewed as symbols, images, propositions, or productions, but they all preserve the meaning of the material encoded.

The fact that we encode material in a code that is like an image was shown by Shepard and Metzler’s experiments. They showed that we manipulate the images just as if they were objects that exist in space.

Production rules are IF, THEN rules that encode our procedural knowledge.

Gentner proposed that knowledge can be thought of as propositions that have predicates and arguments.

Semantic networks consist of nodes (concepts, such as, “vehicle”) and links (relations that express our knowledge about the concepts “vehicles have wheels”)

An early, simple semantic network (Collins and Quillian) was strictly hierarchical. It only had two kinds of links (isa and hasa). It couldn't account for all of the evidence from sentence verification tasks.

A later development (Collins and Loftus) was not – it had many different kinds of links.

It could account for all of the evidence, but it was not a very well constrained theory. You could simply make up a new kind of link to account for any evidence from the sentence verification tasks.

John R. Anderson's ACT model used propositions, arguments, and predicates as the nodes, and had more kinds of links than the Collins and Quillian model, but still only a limited number of kinds of links. It was more constrained, and yet could account for all of the evidence.

Concept learning could occur by a hypothesis testing procedure (for those simple concepts that can be defined by features)

Our natural concepts are not simple “feature defined” concepts. They have a prototype, graded membership, and fuzzy boundaries.