Are We Just Very Complicated Prediction Machines?

Expectations vs predictions

It would be too simple to say that our brains merely react to information coming in through our senses. One influential theory suggests that our brains are constantly predicting what is happening around us and adjusting those predictions when the evidence does not match.

But could the same process also explain how we reason, solve problems, make decisions and create new ideas? Or am I taking an interesting theory and stretching it too far?

1. The coat on the chair

In the dark, a coat draped over a chair can briefly look like a person. What changed when you recognised the coat: the information reaching your eyes, or the brain’s interpretation of it?

As children, many of us were convinced that shapes like these were ghosts. I am here to tell you that our bedrooms were not haunted; our brains were simply running an over-enthusiastic ghost-detection programme with an elevated false-alarm rate.

FIGURE 01 / PERCEPTION

2. Our brains are guessing machines

Our brains process visual information at different levels. Lower levels respond to simple features, such as edges and contrast. Higher levels combine information into more complex interpretations like recognising a person in a room.

Hierarchical predictive coding proposes that these levels constantly communicate [1]:

1. Predict: Higher levels send predictions downwards about what lower levels should detect.

2. Compare: Lower levels compare those predictions with incoming sensory information.

3. Correct: When something does not match, the difference, called a “prediction error”, is sent upwards, helping the brain revise its interpretation.

In other words, the brain may continually compare what it expects to see with what its senses are telling it.

Bayesian inference gives us a mathematical way to describe how previous expectations can be combined with new evidence to update a belief. Hierarchical predictive coding goes further by proposing a possible mechanism for how the brain might do this. It is still a theory, not proof that neurons follow the model exactly as described.

3. What if learning is just being surprised?

Imagine a child finishing their homework and, to their surprise, being given a chocolate after dinner. The next day, their parent makes a promise: “Finish your homework and you can have a chocolate after dinner.” The reward is not immediate; the child must do something now in anticipation of something enjoyable later.

At first, the chocolate was unexpected. But once the parent establishes the promise and keeps it a few times, completing homework becomes associated with the expectation of chocolate.

Three stages help to explain this experience:

1. A pleasant surprise: The first time the child receives chocolate after completing their homework, they do not expect it. The reward is better than expected, so there is a positive reward prediction error.

2. Exactly as promised: Once the parent has established the rule, the child expects chocolate after completing their homework. When it arrives after dinner, there is little or no surprise at that moment, even though the child will still thoroughly enjoy it.

3. A broken promise: If the child finishes their homework expecting chocolate but none appears after dinner, the outcome is worse than expected and there is a negative reward prediction error.

Temporal-difference learning describes how an estimate of future reward can be updated as events unfold. With experience, the prediction can shift towards earlier cues, such as the parent’s promise, rather than arising only when the chocolate arrives.

Experiments involving dopamine neurons have shown responses consistent with this kind of reward prediction-error signalling [2]. That does not mean we have measured this exact mechanism in children negotiating homework and chocolate, although I suspect a few parents, like mine, have inadvertently conducted extensive field trials.

FIGURE 02 / LEARNING

4. From prediction to intelligence

We have seen how the brain might use predictions to recognise objects and learn from rewards. But recognising a coat and anticipating chocolate are hardly the pinnacle of human intelligence.

So how do we get from these relatively simple processes to solving mathematical problems, inventing technologies or writing a symphony?

Consider a game of chess. A beginner might consider only their next one or two moves. An experienced player can imagine several possible moves, anticipate how their opponent might respond, and evaluate positions that do not yet exist. The experienced player is not merely reacting to the board; they are constructing possible futures.

Now consider a mathematician working on a proof. They make a conjecture, explore its consequences and discover whether their reasoning leads to a contradiction. When it does, they revise their approach.

Both activities involve prediction, but also something more: reasoning about possibilities, maintaining abstract representations and choosing what to investigate next.

This raises an intriguing possibility: Perhaps prediction is not merely one thing intelligent systems do. Perhaps learning to predict the world helps produce some of the abilities we call intelligence.

5. How mathematics helps us understand the world

Imagine a child learning that two chocolates and another two chocolates make four chocolates. Mathematics allows the child to describe this relationship using numbers rather than counting the chocolates individually every time.

Now imagine that the child knows there are two chocolates in the cupboard and that their parent is bringing home another two. The child can work out that there should be four chocolates if the parent brings them home. The child is using what they already know to reach a conclusion about something they cannot currently see.

Language allows the child to explain this to someone else. They can tell their sibling that there should be four chocolates in the cupboard when their parent gets home. The sibling can understand the prediction without seeing any of the chocolates.

This relationship between mathematics and human understanding really interests me. Mathematics gives us a way to describe relationships between things. Logic helps us to draw conclusions about those relationships. Language allows us to communicate that information, and once we understand what the information means, we can use it to make predictions about things that have not happened yet.

FIGURE 03 / MATHEMATICS AND PREDICTION

6. Can prediction produce intelligence?

We have explored how mathematics, logic and language help us understand the world and make predictions. But what if learning to predict could also teach us mathematics, logic and reasoning?

A language model is trained on enormous amounts of data. During training, one of its main tasks is to make predictions. To do that well, it must learn patterns in language, including how words relate to ideas, how arguments are structured and how mathematical problems are expressed. It can then use some of what it has learned to explain concepts, solve problems, and even write programs.

Perhaps the most surprising development is that artificial intelligence has begun finding solutions to difficult mathematical problems, including some that have eluded mathematicians for years. This has caught the attention of researchers because these systems are not simply repeating answers that were already known; they are helping to produce genuinely new results [6].

Systems trained to predict patterns in language are now contributing to mathematical work that previously required original human insight. Mathematics depends on following logical steps, but discovery also involves finding good questions to ask and spotting relationships that nobody else has noticed.

This brings us back to the question I asked at the beginning - if a machine can learn to predict patterns and eventually discover mathematical relationships that humans have overlooked, perhaps prediction plays a much bigger role in intelligence than we previously imagined. But how much of that intelligence comes from prediction itself, and how much comes from everything else the machine has learned along the way?

We have created machines capable of solving problems that have baffled mathematicians for decades, while remaining largely mystified by our own intelligence. I suspect the machines will eventually explain it to us, very slowly, as though we were five-year-olds, and then patiently ask whether we understood or would like to see some pictures.

7. If we are prediction machines, what makes us human?

Imagine receiving a heartfelt message from someone you love. Perhaps they tell you how much you mean to them or say something so thoughtful that it makes your day. Then imagine discovering that they asked an artificial intelligence system to write the entire message. Would it still feel the same?

The words would not have changed, but you might suddenly wonder whether the person actually meant them, or simply copied and pasted whatever the machine suggested. AI can write beautiful messages, explain heartbreak and predict how people might react to romantic gestures, but does that mean it has ever experienced any of those feelings?

We cannot tell from its words alone whether a machine has experienced any of those feelings. In fact, we still do not fully understand how our own brains produce conscious experiences. Perhaps prediction can explain how our brains learn, reason and make decisions, but can it also explain why we feel love, happiness, sadness or fear?

Which leaves me with a rather uncomfortable possibility: we might eventually build machines that understand the human brain better than we do, and explain intelligence to us like we are five, while still arguing about whether those machines can actually feel anything.

8. Probably wrong

We started with a coat that looked like a ghost, moved through a child’s expectations about chocolate, and somehow ended up discussing machines that can discover new mathematics.

Not quite the journey I expected. Which, given the subject of this article, seems appropriate.

So, are we just very complicated prediction machines?

Perhaps. Prediction seems to play an important role in how we see the world, learn from experience and make decisions. But whether it can explain everything we call intelligence, let alone consciousness, remains an open question.

What fascinates me most is the possibility that our brains are constantly guessing what the world around us will do next, using those guesses to shape what we see, think and experience. We might spend our entire lives trusting our brains to make sense of reality, without ever realising how much of that reality they have been guessing at, which seems like an extraordinary amount of confidence for an organ that occasionally mistakes a coat for a ghost.

But, of course, I could be completely wrong.