Error Minimization and Action

“Hi, my name is Adam.”

This statement proposes a truth of sorts. It’s not a Platonic one, independent of time. Rather I am proposing the truth of a causal mechanism to you, namely: if you say “Adam,” then I will respond, pay attention to you.

What you are experiencing now is not my individual attention. Still, each of you are capable of predicting what your senses would be should you have my attention: I’d shut up and turn my gaze towards you personally. This is contrary to what your senses are right now, in the present. It is inherently a counterfactual sensory experience.

Yet the existence of this counterfactual sensory experience gives us no alarm at all. We are quite proficient in distinguishing what is a prediction and what is reality. What is real is always much more vivid and available than any prediction.

Nevertheless, our brains are constantly concerned with prediction on many different spatial and temporal scales. Each action we take must be predicted in order to happen. It is not just coarse actions like “I will eat later” that we predict, but each and every movement we take. That’s contrary to how we usually think about acting. It feels like when I do something, that I commanded, willed it to be that way. In fact it is the case that we willed it to be that way, just that predicting it is how we willed it to be.

When we predict some particular sensory experience we create a difference between the two, which we can refer to as error. Our brains want to minimize this error — we have to in order to survive. Imagining a full stomach does not suffice to feed ourselves.

There are then two ways to deal with the difference between our current and predicted senses. The first is to change our prediction to match what is currently being sensed. If I predict that the work fridge contains my lunch and it doesn’t, that error may be reduced by believing some jerk stole it. Now my prediction of a lunchless fridge will match my current experience.

The second is to act to change our senses. Post fridge, I still perceive error between my initial prediction of eating and my current state so I go and get a meal from the cafeteria instead. Once I am in the process of eating, my senses will match my prediction and the error minimized.

The conception of action as changing our senses in order to minimize error, minimize the difference between predicted and current sensory experiences is at the core of what Dr. Karl Friston refers to as Active Inference. I believe that Friston’s work is a seminal insight into how humans and all organisms function.

I gave you a hypothesis at the beginning, and that was: if you experience yourself saying “Adam,” then you will experience my attention shifting to you.

In order to prove this hypothesis we have an intermediate sensory experience that is again not the case, you are not currently experiencing verbalizing my name. So again we must make a prediction: “If I experience myself saying ‘aah,’ then experience myself saying ‘dumb’, then I will will experience Adam’s attention moving.”

We can recurse into this even more, generating a new experiment: “If I experience my tongue in a particular position and experience air flowing out my lungs, then I will experience myself saying ‘aah’,” and so forth.

You’ve experienced saying ‘aah’ so much that there’s almost no uncertainty as to how you bring about that experience. Our brains also represent this precision about how confident we are in a prediction. The more confident about a prediction we are the more surprised we are when the prediction is wrong. The higher the precision the more error is amplified between the predicted and current sensory experience.

Now that you are confident in your prediction on how to change your immediate experience your peripheral nervous system is set up to bring exactly that experience into being.

To recap: say you want me to look at you but I presently am not. When you conceptualize “me looking at you,” you predict what your senses would be to make that true. This creates error between your predicted and current senses. To reduce this error you act. Whereas my eyeballs are under my control, you control proprioceptive sensory experiences like tongue position and airflow. By precisely predicting these experiences over time, you produce the sensation of speaking and shortly after, the sensation of my gaze shifting — despite my gaze not being directly under your control.

Your experiment concludes successfully and your brain learns that a trajectory of some senses, that of speaking my name, will result in other senses changing, your perception of where my gaze is. Your brain changes to represent how it minimized that error, strengthening the connections between neurons that predicted the trajectory of senses. You have learned a way of acting that proves my name is Adam.