Greg Detre
@11 on 4 February 2002
Prof. Rolls, 5 of 8
1-layer perceptron � having target outputs is biologically implausible
will find the LMS error
need to be linearly separable
adaline = linear SLP (Widrow + Hoff) � adaptive linear element
brain doesn't suffer from linear separability issues because huge number of inputs
Rolls says that brain doesn't do logic in the way Minsky + Papert said
PA:������ pmax � C x storage
back projections � don't have 1 input per neuron
(except maybe in the cerebellum)
5% inhibitory vs excitatory
calculating the error is difficult biologically � well, expensive and not found in the brain
bio plausibility
local learning rule
target outputs are a no-no
calculate the error from the target-actual firing
recoding is the trick to solving linearly inseparable problems
the brain doesn't channel through small numbers of hidden neurons
brain has difficulty with discrete-step processing (like in Elman)
settle too fast
neuronal precision is limited � Rolls thinks synapses only really have 8 graded resolutions
brain doesn't like logic, or parity
operates in clamped condition
though there is some synaptic adaptation in forward inputs, allowing recurrent collaterals to settle unclamped
also clamped backprojection
so-called serial performance might result from slow constraint satisfaction
sensory receptor adaptation???
attractor nets are not so good at remembering continuous firing rate distributions � it�s more efficient to effectively binary firing rates
NMDA may do that � only alter the weights if it�s a high firing rate
reinforcement learning
single reward/penalty N for each N/the whole net (Barto & Sutton, 1988)
slow learning
reward vector = just a MLP with noise
useful source of noise in the brain? probably
CS vs UCS???
difference between 1-layer perceptron + PA??? delta rule (supervising target output = error rather than value)
delta rule � learning involves the error rather than subtracting sparseness
LMS error vs �least mean modulus� error???
cable theory??? shunting???
would a 2-layer competitive network be useful???
how do rewards help in unsupervised nets???
www.cns.ox.ac.uk � GAs
trace rule for invariant object recognition???
CA3 � hippocampus???
mossy fibres to CA3 Ns??? colinergic trick input???
RC membranes???
multi-stage PA as AA???