Prediction is intelligence

Much of what we call intelligence involves anticipating what happens next. We learn a pattern, use it to form an expectation, and change our behaviour when the world surprises us.

I find prediction a useful way to connect learning, adaptation and decision-making. Learning gives us a model to work with. Adapting means revising that model when circumstances change. Solving a problem often involves imagining which action is likely to work before taking it.

In neuroscience, predictive coding offers a related account of perception. In Rao and Ballard's model of the visual cortex, higher levels predict activity at lower levels, while differences between predictions and observations help update the model. It is a model of how perception might work; the broader claim that prediction explains all intelligence remains much more ambitious.

Machine learning makes prediction explicit. A system can estimate a future value, classify an observation or learn which actions tend to produce a reward. The quality of those estimates matters. So do the objective and the conditions under which the system will be used.

Consider orienteering. A map suggests a route, but the person reading it must estimate how terrain, distance and weather will affect the journey. A short route can take longer than a long one. Experience improves the estimate, and new information can force a change of plan.

Choosing a career adds another difficulty: deciding what to optimise. A student considering software engineering and environmental science might anticipate income, satisfaction, opportunities and the kind of work each path involves. Even excellent forecasts would leave a choice about how much each outcome matters.

That distinction matters to me. Prediction supplies expectations; judgement determines how to act on them. Two people can agree about the likely consequences and still make different decisions because they value different things or tolerate different risks.

The same reasoning can extend to understanding ourselves. If I can anticipate how I tend to react to criticism, I have a better chance of responding thoughtfully. But the prediction has to remain open to correction. Treating a past reaction as a permanent fact about myself would make it harder to learn.

Prediction earns its value through that repeated contact with reality. I am interested in how well a person or system anticipates events, and in what happens when the anticipation fails. How quickly does the model change? What evidence is enough to change the decision?