Page 418 - Biomimetics : Biologically Inspired Technologies

Bar-Cohen : Biomimetics: Biologically Inspired Technologies DK3163_c016 Final Proof page 404 21.9.2005 11:49pm

404 Biomimetics: Biologically Inspired Technologies

Figure 16.2 Whole body cooperative manipulation of an object.

16.2.2 Motor Learning Using Neural Network

In the learning-based approach, the main efforts have been made through: (1) supervised learning;
and (2) self-organization.
Fundamentally, supervised learning depends closely on the availability of an external teacher. In
this approach, we first construct a neural network and define a smooth nonlinear function for a set of
neurons. Then, for a given set of inputs, we use the error between the desired response from the
teacher and the network’s actual output to adjust the interconnection weights between each neuron.
Researches of supervised learning resulted in the later biological discovery of long-term depression
(LTD) in cerebellum (Rosenblatt, 1962; Ito, 1984), which in turn clarified one of the basic functions
of cerebellum in motor learning and adaptation. However, the later developments of supervised
learning in artificial neural network may not match in detail with the real neural networks
(Rumelhart et al., 1986).
One of the important abilities of supervised learning is the so-called generality, which means
that, after sufficient learning, for a new input that was not learned before, the network can generate
proper output. It is proved for the multi-layered artificial neural networks that, with sufficient
numbers of neurons in the hidden layer, the network can approximate any continuous mapping from
input to output (Funahashi, 1989). For motor learning, however, the condition of sufficient learning
indicates that we have to perform sufficient physical trial motions by the body. This is necessary in
supervised learning but is not efficient for motor learning in biological systems. For motor learning,
the main target is rather to realize the generality of motion with limited physical trials.
By modifying the supervised learning, three models: (1) direct inverse (Kuperstein, 1988); (2)
distal supervised learning (Jordan and Rumelhart, 1992); and (3) feedback error learning (Kawato
et al., 1987; Miyamoto et al., 1988) have been proposed for the specific problem of motor learning.
The main considerations of the modifications are about the selection of the suitable teacher signal
and the concave property of the nonlinear transformation. However, these three models have two
common disadvantages derived more or less from supervised learning. Firstly, in applying an
algorithm such as backpropagation, global information of the network’s output error is used to
adjust all weights between nerve cells. It requires massive connections among all neurons, which is
difficult to realize artificially. Secondly, the resultant motor output may not have topology con-
serving property with respect to the sensory input, or even no spatial optimality as we will show in
the next subsection. Because of these disadvantages, in the tasks such as to move the hand smoothly
in the task space, there may exist a dramatic change in the joint angles (Guez and Ahmad, 1988;
Gorinevsky, 1993).