Physical computation and the design of anticipatory systems

For a system to learn from experience and predict the future, it must have access to the appropriate information. In conventional computational systems engineering, signals of specific types are sought at specific resolutions. Though this is appropriate for many well-understood problems, it may be inappropriate for other, less-well-defined, applications. These include multi-tasking anticipatory systems that interact with the real physical world. Here, the engineer may not have enough advance knowledge to determine what information will be relevant and what will not, particularly in the context of systems that perform complex tasks. Here we discuss how our physical model of computation can allow information to be used more effectively by either postponing, or forgoing, analogue-to-digital conversion. Further - based on evidence from neuromorphic engineering and theoretical computer science - we suggest that systems designed using the approach will be more efficient, both energetically and computationally, than their conventional counterparts.