This is the first of many posts in the new category “Design Decisions”. These posts will chronicle my thought process as I work through various problems, both technical and philosophical.

The first problem that should pop up in any brain simulation project is also one of the most difficult to answer. How does a brain grow? Is the particular structure of the human brain required for intelligence?

Discussion after the jump.

I see two answers to this question:

1. The human brain is intelligent precisely because of its structure. We are intelligent because evolution stumbled on the only correct configuration. You must simulate it exactly to recreate intelligence.
2. The human brain is intelligent because of the parameters governing its subunits. Given a different evolutionary path, the human brain could look radically different and still retain its capacity for intelligence.

Depending on which answer you select, the path towards simulating intelligence is radically different. With option 1) you would begin trying to recreate the structure of the brain, everything from the frontal lobe to the cerebellum, down to the barrels found in the neocortex, down to even the arrangement of pyramidal neurons and various layers. You would have to recreate the various properties of receptors and transmitters, neuromodulators and hormones. You would have to make sure each region functions acurrately to input.

In short, you would have to recreate the entire human brain down to a very precise level. Eventually, given fine enough granularity, you could create an intelligent artificial human brain simply because every aspect of the biological model has been simulated. That is also tens (if not hundreds) of years away.

I have chosen option 2. We are intelligent beings but I highly doubt our intelligence is unique in its structure. Our intelligence arises from a million different evolved traits that coincide synergistically. I would theorize that a million different evolved traits could just as easily create intelligence.

If you take this idea and run with it, you quickly find that this project is not about designing a brain. Rather, it is about designing a system of rules that allow a brain to design itself.

Which brings up the first fundamental design question I encountered: how are connections between neurons decided? One could use a naive algorithm to randomly connect neurons and hope for the best. This will work given enough time….on the order of several million years! A better solution is to give the network a means to create localized structures within the network. The barrels in the neocortex, for example, are dense regions of neurons that have predominantly vertical axon growth and represent a localized structure within a larger structure. It is my goal to give the network the means to create structures like this through the use of a growth system. This should jumpstart the network and give it an advantage that early organisms did not have access to. An important thing to note is that I am not designing these structures, merely giving evolution the tools needed to create whatever structure it sees fit in a shorter timeline that naturally occurring.

I define the growth system as the placement of neurons and with whom they connect to. The growth system in DN must be highly flexible so that the evolution simulation can tweak it to create a wide range of possibilities. It must also be repeatable so that future generations of genotypes can grow similar structures. No two humans have exactly identical number and placement of neurons but they do have identical structures. The growth system must allow fluidity in neuron placement and still retain rigidity in structure. Lastly, it must be granular enough that small structures can be formed. In light of these requirements, I have designed what I call “Growth Hormones”.

Growth Hormones (GHs) are nothing like growth hormones found in the human body. In fact, GHs borrow more from physics in design that biology. Imagine a GH as mathematical point in three-dimensional space. These points generate a field, much like an electromagnetic field. These fields have a force vector and strength which influence the growth direction of axons/dendrites.

To understand how this works, first consider the growth of neurons in DN without GHs. Neurons are placed randomly in a three-dimensional box and then instructed to find connections. The axons and dendrites branch out randomly in all directions from every neuron, creating a fairly uniform solid. The chances of localized structure being formed randomly is negligibly small.

Black lines represent connections between neurons

Now include a handful of GHs scattered throughout the physical network space. Neurons close to GHs have their direction of axon/dendrite growth affected greatly while those further away experience weakened effects. In this way, GHs are the seeds of localized structure. The quantity, position, vector orientation and strength of GH are all manipulated by the evolution simulation.

Black lines represent connections between neurons, green area is the influence of a Growth Hormone

It is theorized that this can provide whatever structure is required for intelligence. More fit structures will be passed onto children genotypes while those less fit will be excluded from the gene pool. This system should be suitably flexible that the evolution simulation can tweak it endlessly. Combined with tweaking of the parameters for neurons, we have an environment that is very flexible. With luck, it will be the appropriate environment for a brain to design itself through evolution.