Jeff Hawkins book, A Thousand Brains, analyzes our latest understanding of the architecture of the brain and attempts to outline a theory of how intelligence works across that architecture.
Hawkins initially skirts Deutsch's fundamental philosophical and epistemological questions of how humans are actually able to create explanatory knowledge and instead focuses on a model of how the neocortex works. Later in the book he attempts to address some of these questions and related quandaries, with what I would personally consider inadequate arguments and definitions. That being said his approach is still useful in that it provides an overall explanation that can be critiqued.
He describes the neocortex as an evolutionarily recent addition to the mammalian brain that is responsible for our "intelligence". The neocortex is wrapped around and deeply interconnected with an older "animal brain" that controls our autonomic systems and instinctual motivations/emotions. The neocortex is remarkably uniform in its structure, being composed of similarly structured cortical columns that vary in only minor ways across the whole neocortex. The cortical columns are filled with neurons that process detailed sensory information flow. The relatively similar structure of cortical columns across different sensory regions of the brain supports the concept of brain plasticity.
The first part of the book describes a model of how information is processed across the neocortex. In a nutshell it comes down to several key concepts that revolve around the idea that: The neocortex learns a predictive model of the world.
Key points:
Cortical columns learn models of objects
These models include reference frames
Reference frames are related to grid cells that evolved in our older brain to help animals map/move about in their environment
In my view reference frames could be crudely equated to knowledge (or existing explanations of what we are sensing)
Movement creates incoming sensory signals that update our model (explanation) relative to these reference frames. In any given situation we often expect to see, feel, or hear something, and if we don't our predicted model (explanation) needs to be updated.
This process evolved from physical movement but does not necessarily require it in humans. Maybe this is why humans are universal explainers and animals are not? - we can think without moving.
When information comes in that does not fit the model, what happens?
Voting! - to me this is very interesting and has parallels with Deutsch's view of how knowledge is created through conjecture and criticism. How does voting work?
Most information flow in a cortical column stays within the neurons of the column and represent detailed (lower level) sensory inputs, but some cells in the column connect outwards and allow higher level (emergent) information representing an explanation like "coffee cup" or "democracy" to flow across the brain and to other columns.
When conflicting information arrives in a cortical columns that does not fit the predicted model (pre-existing explanatory knowledge) the model has to be updated, and each column comes up with a new model or explanation.
Cortical columns contain different inputs and often arrive at different models (explanations) but through a networked system of voting we converge our attention on the model (explanation) that is most common.
Attention is an interesting part of the puzzle that must be an important for any intelligence algorithm. Conflicting models (explanations) in our brain can't always be reconciled through voting, and by shifting our attention we can choose one or the other. An example from the book is shown here, you can see a vase or two faces, but not both at the same time.
There is so much more detail in the book, but what I've described is intriguing to me as a possible pathway to uncovering the algorithm behind our explanatory creativity. Hawkins' approach is more of a reverse engineering of the brain, but he has an overall explanation that we can discuss, criticize, and learn from.
Hawkins doesn't address important epistemological questions adequately, but perhaps answers to these questions will fall out of further discoveries that come from trying to program this type of brain architecture.