Plaut and MLPs
- Remember MLPs? While they were useful for some machine learning
applications in the 80s (and still are), they were also used
for cognitive models.
- A good example is Understanding Normal and Impaired Word Reading:
Computational Principles in Quasi-Regular Domains, Pluat et al (1996)
- It's about pronouncing words from the spelling. It pronounces
all the words it's trained on correctly, does both pronounciations
of ambigiuous words (like read), is slower to learn irregular words, etc.
- Now an MLP is not really much like a brain, so how can
Psychogical phenomena emerge from them?
- There are two keys here. The first is distribution. Here the
knowledge the system has is distributed over, to some extent,
all the Perceptrons in the net.
- The second is learning. The system actually learns to solve the
problem.
- So, it's a nice model, though it's not a neural representation of
how people do it.
- Now I'm seeing papers about how deep nets represent parts of
cognition (and weaker papers that say all of cognition).
