In en earlier post I defined knowledge as useful information. I used the word “useful” instead of “true” because we can never know anything for certain, even if some theories like quantum electrodynamics yield predictions with an accuracy of twelve significant figures.
Usefulness can be defined as the capacity of information to contribute to the well-being of an organism. In biological systems, well-being corresponds to maintaining adaptive states that ensure survival and reproduction. In the active inference framework, useful information minimizes free energy by enabling the organism to predict and control its environment effectively.
The core of the active inference framework is a model realized in the brain (or some simpler control system) that (1) correlates actions with (subsequent) mental states, (2) correlates observations with mental states, and (3) stores experiences in the form of prior probabilities (prior making any observations) of possible mental states. (In simple organisms the mental states may be the same as, or very close to, the observations.) The organism’s desires are encoded either in the priors (if a simple organism) or in a separate model. The model is in AIF represented by a mathematical statistical model.
Organisms use the model both to interpret their observations and to predict what mental states they will end up in given certain actions or series of actions. The organism then use the information thus obtained to select a course of actions (policy) that leads to desirable mental states (and thus desirable observations).
Organisms don’t have direct access to the real-world states. They only have access to the world through their senses and, in the case of humans, through measuring devices that measure some limited aspects of the real-world states. This is the reason for why they can never be certain that they know everything there is to know about a real-world state.
For the model to be useful, the mental states inferred from our observations need to be aligned in a meaningful way with real-world states. If the organism encounters a real-world predator, then its mental state should represent a predator with a high probability. If it feels hunger, then the feeling should correspond to real need for food. Real-world states and mental states get aligned through evolution and through learning during the life-time of the organism. Evolution has given us humans a model adapted to life on the savannah but didn’t give us a mental state for ionizing radiation. It had to be learned by accident and later through science.
If we have a good correspondence between mental states and real-world states, then the model introduced above is a good model of the world. Given that the main use of knowledge is to navigate through the world and through life, we can say that ontologically knowledge is our model of the world.
It is also notable that the model is a “forward model”. It doesn’t render mental states from observations but (predicted) observations mental states. The brain has to iterate to find the mental state matching the actual observation to the predicted observation. Likewise, the model doesn’t render optimal actions from desired mental states but the brain has to run “simulation” (or use some other iterative methods) to test out different actions to determine which one will lead to (near) a desired state. Since evolution has chosen such a solution, the forward model must have some significant advantages even if the instinct of an engineer would be to try to find the inverse models that would directly give mental states from observations and actions from desired states.
The fact that the models of AIF are “forward models” does align nicely with the claim that knowledge (useful information) can be seen as a model of the world.
Epilogue
I also realize I now need to replace the quote from the Stanford Encyclopedia of Philosophy on my home page:
Most epistemologists have found it overwhelmingly plausible that what is false cannot be known.
With my definition of knowledge above, you can in fact know something that is false. Since strictly speaking, most information is to some degree false.