Information & Control Versus Computation
A recent book, Possible Minds (Brockman, 2019), provides 25 essays on the future of AI, building upon Norbert Wiener’s 1948 classic Cybernetics: Control and Communication in the Animal and the Machine. A key distinction among these pundits is between information and control versus computation.
This distinction is intriguing. My roots are definitely in the information and control camp. Phenomena such as dynamic response, feedback control, state estimation, and response stability have long been central to my thinking. How is the computational view different? One distinction is continuous vs. discrete time systems, but this cannot be the essential difference.
Perhaps it is rule-based behavior, either symbolic logic or statistical, as opposed to equation-based behavior. I think that the difference is more subtle. Wiener assumed that control and communications are based on humans adapting to the natural phenomena being controlled, as described by differential equations and stochastic relationships.
An interesting instance of this view is the notion that humans are “constrained optimal” controllers and decision makers. They perform as best as possible within the constraints affecting their performance. For example, consider manual control of vehicles or processes. Humans have neuromotor lags (i.e., they cannot move their limbs instantly) and their observations and movement are “noisy” (i.e., sensing and movement are not precise). Solving the optimal control problem with these constraints often provides accurate predictions of human performance.
Constrained optimality works best when people have no choice but conform to the dynamics of their environment. There are many situations, however, where people have considerable discretion both in what they do and when they do it. Our studies of troubleshooting performance provide a good example. People’s choices and timing of tests varied considerably. We found that rules based on symbolic logic could predict troubleshooters’ behaviors. This included rules that were recalled, applicable, useful, and simple. Hence, people’s experiences influenced their rule sets, not just the physics of the failures of interest.
This view is based more on computational representations of humans’ rule sets and how rules are accessed than on information and control theoretic dictums. Tasks involving recognition and classification of visual and aural patterns seem to also be better represented computationally. Recent successes in machine learning suggest that computational models based on multi-layered statistical representations provide support for this view.
There have been criticisms that such models cannot explain themselves and, therefore, their applicability to decision support may be limited. However, there are many things that cannot be directly explained. If presented with a picture of someone you identify as a family member, friend, or pet and asked to explain why you think it is them, it would be very difficult to provide a definitive explanation.
Nevertheless, the idea that pattern recognition and classification is computational seems difficult to dispute. In the early 1970s, I took Marvin Minsky and Seymour Papert’s course at MIT on computer vision. Their lectures included both rule-based and statistically based approaches. It was amazing how many rules were needed to simply recognize that something was an edge. Symbolic logic was not a viable approach to the general computer vision problem.
So where does this leave us relative to the distinction between information and control versus computation? It seems to me that neither view prevails. When humans have to comply with the dynamics of their task environment, information and control works well. When humans have considerable discretion, computational representations are better, using symbolic logic when there are clear rules of the game and statistical models when experience can be accumulated over many instances of the phenomena of interest.
Of course, we need to keep in mind George Box’s aphorism, “All models are wrong, but some are useful.” So, the ultimate question concerns what a particular representation allows us to do. What questions does it help us to address? What types of predictions does it enable us to confidently make?
