Let's escapes boredom with Andy Clark

This post is based on A nice surprise? Predictive processing and the active pursuit of novelty (2018), Andy Clark.

The Brain is a Prediction Machine (Predictive Processing Theory)

Basic idea: the brain is made for predicting what it would perceive. And experiments show how it is using Bayesian inference.

There are two ways that the brain can satisfy its purpose of predicting what it senses. One way is to improve its own prediction algorithm, to understand the outside world better. This is learning. Another way is to mess around the outside world, to make it so that its predictions come true. This is acting.

It is quite amusing to think about action in the sense of prediction:

1. The brain predicts that the hand is going up.
2. The input from the eyes show that the hand isn't up.
3. This creates a big error term that the brain really tries to minimize.
4. So the motor cortex messages the arm muscles to contract and bring the hand up.
5. The hand is going up.

Death in a dark room

The obvious problem with this idea is that such a brain would live a really boring life. Basically, a brain that tries to predict the future could do that by living a really boring life, but humans actively seek out non-boring lives. This is the "dark room" problem.

If brains like ours are driven only to minimize prediction error signals, why don’t creatures like us simply find a nice dark corner (providing fully predictable, meager, unvarying patterns of sensory stimulation) and stay there, slowly growing weaker and then dying?

There have been a few ideas for solution. One possible solution is by assuming that evolution has given us a prior in the brain preinstalled, so that we expect to find a world that's distinctively not a dark room. Then, we would avoid dark rooms, because we predict ourselves to not be in the dark room.

This answer is dubious though, since pretty much any activity can be justified with enough manipulation of the prior, and that'd make the theory unfalsifiable.

A better answer is that dying is not predictable, actually. As we die, our body decays and becomes a hot mess, the very opposite of predictable. So, in order to keep our ability to predict our own body, we must keep our own bodies in homeostasis. This is the so called "interoceptive predictive processing".

A boring life

Still, out of the dark room, one might fall into the "boring life". Why wouldn't a prediction brain merely do as little as possible to survive?

long-term minimization of prediction-error might mandate ‘stereotypic self-stimulation, catatonic withdrawal from the world, and autistic withdrawal from others’.

Again, we can assume there's some kind of preinstalled prior in the brain that predicts some exploration. Also, note that some people are more interested in play while others are less, this is probably an evolutionarily stable strategy, to have a mixed population, to deal with a wide variety of environments.

A Mauderate Life

But even this isn't enough. There's still the Merely Modest Exploration Trap. Basically, a prediction machine would not be truly creative or seek true novelty. It would do nothing more than to maximize its survival and homeostasis.

These prediction error minimizing agents remain locked, it seems, into an information-theoretic journey whose guiding principle is in some way unacceptably conservative. It is a journey which, if successful, will be marked only by the attainment of expected goals and meta-goals... Doesn’t this leave unexplained the basic human need for ‘self-actualization and personal growth’?... For such agents, the ultimate information-theoretic goal is a state in which there is zero prediction error.

A Mauderate Life

The current guess is that for some reason, a lot of humans have started seeking novelty for the sake of novelty. It's like there's some kind of meta-prediction, a prediction that says "I'm curious and will keep being surprised".

However, this seems also a problem with this, since any curiosity seems to have a "weak spot" that can get "trapped" in some environment that's not "truly curious". Like this AI that got addicted to watching TV.

I think the answer would be some kind of self-referential stuff like "I will surprise myself.". So that whenever you found yourself able to predict yourself in some way, you immediately look for ways to make yourself not predictable in that way.

But immediately that fails. I am alive. I expect to stay alive for years. Should I immediately look for ways to Surprise! die? Nah. Similarly, there are plenty of things (like sleeping when tired, eating when hungry) that are always pleasant no matter how many times it's done, how predictable it gets.

So the commandment to always surprise myself is definitely not how humans work. What humans seem to do is to surprise themselves in a few narrow domains and call it a day.

For example, it's been a few thousand years, but most human stories fall into about a hundred tropes and shows no sign of quitting them (in fact, any story that deviates from them is unpopular and forgotten). As another example, falling in love is way worn out, but people still do that anyway. In general, despite the human desire to "surprise themselves", human culture still stay within certain homeostases and cannot leave them, not unless humans become no longer human.