- A clone hive of neurons. As each neuron is a clone, having the same set of genes as all the others, selfish gene evolution results in the neurons acting altuistically toward each other. The neurons display a mixture of cooperative and friendly competitive behaviour. As in an ant hive, there are specialized castes of neurons, eg dopaminergic, serotinergic. Castes can be identified by the neurotransmitters they emit. Neuron-neuron interactions occur on the basis of these neurotransmitters: signal transmission, axon and dendrite growth and pruning, synapse reconfiguration.
- Approximate. A guess: A redundant structure, with important components replicated by many neurons (n=log importance?), these neurons in friendly competition to provide the best service. A graceful degredation curve for less important components (eg memories, skills). That is, less important structures are more affected by accidents of brain architecture (eg mapping to a plane -- self organizing maps). More important structures more accurately obey some abstract mathematical scheme.
- A novelty seeking sampler. The sampler seeks information that will tend to rapidly compact the possible-world-model probability distribution -- this seeking might consist of thought (eg dreaming), or actions. The process is random in nature, like a Metropolis sampler.
- The key feature of this is the ability to make good diagonal leaps. This is the major component of "intelligence"/"talent". When we work out how to make computer do this, they will be recognizably intelligent.
- The active set of parameters (short term memory) is quite small. Sufficient to locate it within a larger dormant structure, but not a significant proportion of that overall structure. I.e. the size of short term memory is approximately the log of the overall brain size.
- The sampler must operate over the whole structure, not just the active set. It may do so slowly. The dormant portions probably contain not one but several or many samples.
- There is something I am missing here.
- The key feature of this is the ability to make good diagonal leaps. This is the major component of "intelligence"/"talent". When we work out how to make computer do this, they will be recognizably intelligent.
- A utility optimizer, using the sampler to perform certain integrations necessary for utility calculations. This is the "animal", the "beast within", the "id". The utility optimizer shares control of the body with the random sampler, neither is dominant. Dopaminergic neurons? (This seems to have progressively shut down in my own brain over the last few years ?)
- A series of moments of conscious awareness. A tendency to altruistic action that has no obvious reason to have evolved (from a "selfish gene" perspective). This tendency fades as the "framerate" increases.
Addendum regarding sampling:
- The possibilities being sampled are not terribly complex. Brute force look-up tables of experience rule, not elegant recursive abstractions. Most of the computational load arises because there are so many simple possibilities.
- There is no "sub-conscious", though there will always be bits of the brain that aren't actively contributing right now (the majority of it, in fact). You are conscious of most of what is going on in your brain, and what is going on is usually not terribly complicated.