BBS Commentary on : Andy Clark , ‘ Whatever Next ?

Although predictive coding may offer a computational principle that unifies perception and action, states with different directions of fit are involved (with indicative and imperative contents, respectively). Predictive states are adjusted to fit the world in the course of perception, but in the case of action the corresponding states act as a fixed target towards which the agent adjusts the world. Main Text One of the central insights motivating Clark’s interest in the potential for predictive coding to provide a unifying computational principle is the fact that it can be the basis of effective algorithms in both the perceptual and motor domains (Eliasmith 2007, 380). That is surprising because perceptual inference in natural settings is based on a rich series of sensory inputs at all times, whereas a natural motor control task only specifies a final outcome. Many variations in the trajectory are irrelevant to achieving the final goal (Todorov and Jordan 2002), a redundancy that is absent from the perceptual inference problem. Despite this disanalogy, the two tasks are instances of the same general mathematical problem (Todorov 2006). Clark emphasises the “deep unity” between the two problems, which is justified but might serve to obscure an important difference. In the perceptual task a prediction error is used to change expectations so as to match the input whereas, as Clark notes, in the motor task the prediction error is used to drive motor behaviour that changes the input. In perception, prediction error is minimised by changing something internal (expectations) whereas in action prediction error is minimised by changing something external (acting on the world so as to alter sensory input). Although it is true in one sense that there is a common computational principle