A special application for neural networks are internal models, which relate motor commands to the sensory change that they cause. Two types can be distinguished (figure 1.2). First, a forward model predicts the effect of an action; that is, it maps the current sensory state St (at the time t) and motor command Mt onto the sensory state at the next time step, St+1. Second, an inverse model computes the motor command Mt required to change the sensory state toward a desired value St+1.
In human physiology, internal models were suggested to be an integral part of motor control (Kawato et al., 1987; Wolpert et al., 1995). For goal-directed movements, inverse models are necessary to directly compute motor commands. On the other hand, forward models can predict an outcome of an action before the sensory feedback (via the environment) is available. Moreover, they might be used to cancel the sensory effects of movements (Blakemore et al., 2000) and to predict the consequences of actions without overtly executing them (Wolpert et al., 1995). Chapter 7 will use a forward model for prediction.