17.01.2013 Views

Chapter 2. Prehension

Chapter 2. Prehension

Chapter 2. Prehension

SHOW MORE
SHOW LESS

You also want an ePaper? Increase the reach of your titles

YUMPU automatically turns print PDFs into web optimized ePapers that Google loves.

Appendix C - Computational Neural Modelling 393<br />

4 Generalized<br />

Figure C.6 The generalized delta rule is shown, where a teacher is<br />

needed to compute the error between the desired output and actual<br />

output.<br />

Awij = (tpi - *i> 9j (13)<br />

where the actual output Opi is subtracted from the desired output tpi for<br />

a particular inputloutput pair, p. For the output layer of a layered<br />

network, it is the same as the delta rule. For the other layers, the error,<br />

E, is still dependent on the weights, but a continuation of the chain<br />

rule must be applied to show the mathematical dependence. The error<br />

is propagated backwards.<br />

To perform a computation in an adaptive neural network using the<br />

generalized delta rule, two phases must occur. At the start, weights are<br />

initially random, or arbitrary, values. The training set is presented to<br />

the network, one input/output pair p at a time. The first phase is a<br />

forward computation using equation (3) to sum the products of the<br />

inputs and weights to each neuron in order to produce an activation<br />

value. Equation (4) is used to threshold the result. This computation<br />

propagates forward for all neurons in the network. The results at the<br />

output neurons are then compared to the desired outputs stored in the

Hooray! Your file is uploaded and ready to be published.

Saved successfully!

Ooh no, something went wrong!