Lecture 16 - Debate - Edward Rolls VS Kim Plunkett
Biologically Plausible Neworks - Edmond Rolls
Local Learning Rule
Rarely a seperate "teacher" for each neuron
Pattern association is likely
No multilayer backpropagation net
No information processing with limited numbers of hidden
neurons to improve generalization
Uses large number of neurons and expansion recoding to
simplify subsequent stages.
Though there are back-projections in the cerebellum, they
don't seem to propagate error
The brain is a terrible discrete-step processor (seems to be
continous dynamic of the "integrate and fire" neurons).
Neuronal precision is limited only 4-5 distinct firing rates.
The brain does not solve logic equally (fixing XOR may not be
neccessary)
Connectionist models are helpful in providing insight, but an
aim should be toward ever more biological plausibility.
Connectionist Nets - Kim Plunkett
Need to let many flowers blossom, even symbolic accounts. So
the question is: Is the intermediate level provided by backprop
really helpful to that endeavor?
Well, whatever you do, you simplify. Simplification is a
necessary part of any model building. So don't think that there is
any good reason to throw out backprop just because the evidence
isn't in. Particularly, because we seem to get the same results
anyway. Backprop is part of a class of algorithms, of which
Hebbian is a part. Thus, it shares many similar features. Why not
use it?
To look at the kind of problems connectionists like to study
(like language, cognition, etc), we really have no idea about the
real networks involved. Indeed, the level of specificity we use is
actually an abstract representation of the space, using known
computational skills of the brain. Thus, as we can bring to bear
actual networks involved in language, by all means use them.
But, until then, such abstract accounts may actually inform
upon empirical research.
Finally, the issue of teacher signals is often misconstrued.
Connectionist networks are usually drivin by positive evidence
with an internally generated error signal.