How do we use the Connectome for AGI?

Using the connectome for AGI is the tricky part of the process. Tricky because it does take some thought behind how to associate sensory neurons to real world concepts and how to read the motor output in order to determine the results. To do this we break down the process simply as Sensory Input > Cortical Processing > Motor Output (> Sensory Processing > Cortical Processing > ...).

Sensory Processing

Each connectome has a set of sensory neurons. When we create our AGI, we connect weights to the sensory neurons based on how we perceive the sensory input on a particular sensory neuron. As an example, if we want our AGI system to avoid particular situations, we connect that sensory input to sensory neurons that would cause the connectome to sense pain or unpleasant conditions. If we want our AGI system to want or like a particular input, we connect that sensory input to sensory neurons that would cause pleasant conditions in the connectome. Using Robotics makes this process much easier because it is mostly obvious how we want our robot to react to certain types of stimuli. We want to avoid the robot crashing into an object so we would want a sonar sensor to stimulate sensory neurons that a biological counterpart would sense as painful. 

In my C elegans experiments, I set the sonar sensor to stimulate nose touch sensory neurons and I used sound at a certain decibel to stimulate sensory neurons associated to the presence of food. The stimulation of the nose touch neurons caused the robot to avoid objects by stopping and moving back/around the object but upon my whistling, it would move forward for food. 

One of the advantages we have using emulations of biological brains is we can add sensors well beyond our biological counterparts. For example, one of the things I've wanted to create is a prospecting robot where a metal (gold!) sensor can be used to stimulate a pleasure area of the brain so the robot will want to seek out and find precious metals, marking where those metals are found. Whereas a dog might mark its territory by peeing in that area, we can trigger a similar behavior by "peeing" a powder dye in the area that the metal is found.

Another example that may not be as obvious as what we can do with robotics. We may want to find the best deal on a car. I could create an hourly web scrape to view pricing at certain dealerships and when the price goes down, I stimulate pleasant feeling sensory neurons, and when a price goes up, I stimulate unpleasant feeling sensory neurons. In addition, it is imperative to add multiple sensory feeds to the connectome. The correlations within the connectome/cortical processing is key to AGI and one of the major differences between AGI and current Deep Learning approaches are DL is singular sensory in nature. In my car pricing example above, maybe multiple car dealerships would make this possible or maybe we add some commodities like the price of steel and aluminum? Or maybe the price changes between the Euro and the Yaun? The latter may seem unrelated but the cortical processing may surprise one on how seemingly unrelated items can in fact have a correlation.

Cortical Processing

The connectome is the processor. It reads sensory information and from that determines the output which is the physical or imaginary movement of muscles. There are two types of cortical processors based on the connectomics. The first is what I refer as static connectomics. An example of a static connectome is C elegans. C elegans does not have plasticity in the common sense. We know it does through changes in neurotransmitters but for the most part the synapses are unchanged. This type of connectome is simple and can be manipulated through sensory variations. The muscle output is fairly well defined based on the sensory inputs.

The second type is what I call dynamic connectomes. This is where plasticity is at play and the strength, weakening, new growth or synaptic pruning happens. The network literally alters in accordance to sensory-motor interactions or as I feel is a better way to express it, motor-sensory interactions. Everyone following my work knows that I believe motor skills are the key to cortical adaptation. Sensory input is just the means for us to sharpen our motor output.

Motor Processing

We have a brain for one reason and one reason only - that's to produce adaptable and complex movements. Movement is the only way we have affecting the world around us ... I believe that to understand movement is to understand the whole brain. And therefore, it's important to remember when you are studying memory, cognition, sensory processing, they're there for a reason, and that reason is action. 

                       -- Daniel Wolpert

Motor processing is probably more difficult to understand than Sensory or Cortical processing because it is a constant flow of activity. It's not like we have a result that is constant. We have many results flowing to us over a function of time. 

In my experiments, I see constant multiple motor neuron stimulation of muscles. The very interesting observation I see is that a lot of that motor stimulation is ignored in the sense that it's too weak to cause a contraction. Like in the cortex, we have a lot of neuronal stimulation, but only certain stimulation makes the changes we want to experience a thought. The rest are wasted for the moment although that minor stimulation can turn into the major stimulated network on the next thought, primed by the earlier minor stimulation.

In addition, I find that the minor stimulation of muscles, although too weak to cause a major contraction, is what gives us fine motor skills. Most robots today, despite so many attempts to create smooth movements, are jerky when we get right to it. When our motor cortex sends contraction signals, it stimulates a lot of muscle cells, some stimulation is direct and very forceful while there are many other muscle cells being weakly stimulated, creating a concert of action with fine motor skill as the result.

When you use connectomes as a basis of action, one must always note: 1) the connectome (brain) is always on, and thus 2) motor neurons are always being stimulated. This can be challenging in the sense that applying this to a robot means that the robot naturally will always want to be moving. Has one ever seen C elegans kicking back and relaxing? No, it doesn't happen. It may slow down when it comes across an area of food but what is happening is that eat/digest muscles are now active more than body muscles. An old backpacker's trick is to eat cheese or candy before going to bed in cold climates because even though you are asleep and not walking, your stomach muscles are working overtime trying to digest difficult foods and thus keeping you warm at night. However, we all know, we move quite a bit in our sleep. 

Unlike Deep Learning et al, where we enter an input and get a resulting output, our nervous system is creating output constantly to better affect our environment and creating a true AGI robot is difficult dealing with continual movement. I generally turn an AGI robot on and let it go. With AGI, you don't turn on the robot and then give it commands to step though; you turn it on, let it go and then using sensory stimulation, affect the action in a manner you want the robot to behave - hmmm, just like animals :-)