Chapter 2. Prehension

86 THE PHASES OF PREHENSION
(surface length and object width) and two task properties (magnitude
of anticipated forces and precision needed). The network computed
which opposition should be used (either pad or palm opposition),
given the object and task properties. Simply stated, the neural net-
work learned how to choose between palm and pad oppositions given
the object characteristics of length and width and the task requirements
for force and precision.
An adaptive multilayered network of simulated neurons was con-
structed for doing this computation (see Figure 4.8a). The network
consisted of four input units (bottom row of neurons), four hidden
units (middle row) and one output unit (top row). An iterative learn-
ing process called supervised learning was used to train the network.
A given task (surface length, object width, amount of force, and task
precision) was presented to the input layer. An opposition was chosen
by summing up weighted activation values of the input units and then
weighted activation values on the hidden layer. If any of these
weights were set wrong, as they will be initially, the computation will
be inaccurate. This computed mapping was then compared to the
desired mapping (in the training set). If there was a difference, Iberall
used the generalized delta rule to adjust the weights between the input
units, hidden layer, and output units in order to reduce this difference.
The training set was drawn from the data points shown in Table
4.1. These were compiled from experimental (Marteniuk et al., 1987)
and observational data that pertains to mapping task requirements to
grasp postures. Input requirements are stated in terms relative to the
hand’s size and force capabilities. Object size has been broken down
into two separate properties: object length and object width. Length is
measured in terms of the number of real fingers that can fit along the
surface length at the opposition vector, and width is in terms relative to
hand opening size. Task requirements have been broadly and subjec-
tively characterized in terms of power and precision requirements:
magnitude of forces (without regard to their direction) and precision
needed. Grasp postures are characterized by the chosen opposition.
In the case of grasping a beer mug by its body, the surface length is
greater than four fingers, the width of the mug is generally large rela-
tive to the span of the hand, and the weight of the filled mug is rela-
tively heavy. Lifting it requires a large translation upward, involving
little precision. In this task, palm opposition has been observed. For
heavy objects, palm opposition tends to be used with as many fingers
as possible. As the weight of the object decreases, the posture
switches to pad opposition.
For computing the difference between the desired output and the