Robotics Reference
In-Depth Information
what are called the
strengths
of these connections between the neurons.
25
Setting appropriate weights is the really clever part—an ANN usually
learns its weights from examples. If the task is to distinguish the spoken
digits from “zero” to “nine”, the ANN will normally be presented with
dozens or even hundreds of samples of different people speaking those
words. It is rather like saying to the ANN: “I am going to play you
recordings of 500 people saying the word 'zero' so that you can learn
what the word sounds like.” The ANN then learns from the experience
of hearing all these recordings, repeatedly adjusting its own weightings
until it has learned as much as it can from the sample data.
An ANN normally starts life with random values assigned to each of
its weights, so at that stage the ANN knows absolutely nothing. The
network is then exposed to what is called the training set of data, such
as the 500 samples of the spoken word “zero” referred to in the previ-
ous paragraph. When the first training pattern (also known as an input
pattern) is presented to the network, the weights on each of the con-
nections are adjusted by the program itself by a very small amount, in
order to increase the likelihood of the network recognizing that pattern
if it sees it again. Looked at in numerical terms, the changes to these
weights will make it more likely that the sum totals of the various calcu-
lations at the artificial neurons will result in those totals classifying that
same input pattern correctly. Once this has been done for the first input
pattern in the training set, a second pattern is presented to the network
and the process repeated, and then again for each of the patterns in the
training set. When every pattern in the training set has been presented
to the ANN, the whole process is run again and again with all of the
training patterns, hundreds or thousands of times or more. The reason
the weights on the connections are only adjusted a very small amount
each time is to allow the ANN to learn to recognize
all
of the patterns
reasonably well. If the adjustments were larger, the ANN would be able
to recognize the most recently presented pattern every time, but all the
other patterns never.
The Creativity Machine
A remarkable and highly original application for ANNs has been devised
by Steve Thaler, a physicist turned AI-researcher. During the late 1980s
25
The strengths may be thought of as measures of how important it is, in a particular classification
task, that some combination of two or more different inputs are present simultaneously.
