Shepard Cognitive Model
- Shepard, Hovland and Jenkins have a classic paper from 1961,
Learning and Memorization of Classifications.
- They take eight items defined by three binary features:
Squares and Triangles, Black and White, and Small (Ickle) and Large
- They note that there are six types of binary categories that you can
form with four in each category.
- The simplest (type 1) is based on one feature (e.g.) Triangles vs.
Squares.
- The next (type 2) is based on two features (e.g.) category 1 has
Black Squares and White Triangles.
- People categorise type 1 best, then type 2, then types 3-5, then
type 6.
- I thought it would be easy to plug in our
categorisation system .
- I plugged it in and it did well (perfectly) on type 1 but not the
others.
- This comes from the xor problem. Each feature contributes
equally to each category, so it's hard to get a benefit.
- We changed the input so that instead of binary inputs it took
three tuples as inputs, e.g. IBS. This did perfectly on all the
categories. It was learning eight CAs.
- That's not a good cognitive model.
- So, I used three types of input features, binary features, pairs
and triples.
- This hasn't been worked through and isn't published yet, but gets
close to the right results.
- It's still unsatisfying. People can't have these three tuples
hard coded can they.
