Error-Driven Task Learning - Computational Explorations in Cognitive Neuroscience - page 170

culation algorithm. An improved form of GeneRec is
equivalent to the contrastive Hebbian learning (CHL)
algorithm, which we use in our networks.
For biological synapses to compute the weight
changes necessary for the CHL algorithm, they would
have to perform LTD when the expectation state for
a unit exceeded its outcome state, and LTP otherwise.
Assuming a relatively rapid transition between the ex-
pectation and outcome activation phases, one would ex-
pect LTD with a transient (erroneous) expectation, as is
consistent with the biology of LTD. The GeneRec algo-
rithm thus provides a biologically plausible yet power-
ful form of learning that can learn arbitrary input/output
mappings (tasks).

5.11 Further Reading

The Chauvin and Rumelhart (1995) topic on backprop-
agation is a good source for basic ideas and further de-
velopments in backpropagation learning.
Crick (1989) provides an influential critique about
the biological plausibility of backpropagation.
For further details on the GeneRec algorithm, consult
O'Reilly (1996a) and O'Reilly (in press).
The journal Neural Computation and the NIPS con-
ference proceedings ( Advances in Neural Information
Processing ) always have a large number of high-quality
articles on computational and biological approaches to
learning.