Poor NeuroCognitive Model
- As a neurocognivie model, it has some real flaws.
- Firstly, it is a supervised task. Humans clearly don't learn
these things in a supervised way.
- The training and test sets are really small.
- The volume problem is just ignored.
- The topology is unrealistic.
- The hardcoded CAs aren't realisic, though probably easily fixable.
- A1 doesn't actually work like this (see Ohl and Scheich 1997).
- No backward connections.
- There are many missing layers.
- Obviously, it only deals with the easiest vowels and should deal
with a larger or full range.
- Also, the learning portion of the task itself is not a good
neurocognitive task; people (and presumably gerbils) don't learn
from two instances of three sounds and generalise to 10.
- Similarly, it doesn't deal with a range of other hearing problems
(e.g. consonants).
- Despite these (really serious) shortcomings, it's the best
neuro-cognitive model I'm aware of, simply because it takes
input as a human does.
- That means that it's not hard to do better. Hopefully some of you
will give it a go, and I'll probably get back to it.
