Recurrence

One of the surprising revelations that I discovered while preforming emulations of animal connectomes is the huge amount of recurrence. The brain literally folds back on itself over and over again. 

To illustrate this idea, let's look at C elegans again. Keep in mind that we are working with 302 neurons. Neurons connect to a set of other neurons. This connected set has a number of axons that link back to the originating neuron. If we look at the Interneuron AVAL, AVAL connects to 77 other neurons. These 77 neurons connect to another set of neurons, so on and so on. Each neuron can have multiple connections to other neurons. For example, AVAL has 11 synaptic connections to the neuron DA6. From the first set of neurons that AVAL connects too, there are 83 total synaptic and gap junction connections from those 77 neurons back to AVAL. If we take those 77 neurons and follow what neurons they connect too, we find from that second set (Layer 2), there are 573 neuronal connections back to AVAL. If we just go out 5 connections deep, we have over 2 million connections back to AVAL. 

Recurrence is exponential in a nervous system

The chart below shows a number of interneurons and their recurrent connectivity.

We most often think of artificial intelligence systems as feed forward systems. Yes, we have AI that has feedback loops for weight variance and reinforcement learning but nothing to the scale of the animal brain. I believe to create AGI, we must incorporate the concept of exponential feedback loops. 

As noted in an earlier post, our brains are always on. This is why. When I started the C elegans connectome emulation and later wanted to stop the individual programs that made up each of the 302 neurons, I would find quite quickly there were sets that would continually be firing to one another. Neuron X fires to Neuron Y and Neuron Y meets it threshold very quickly and fires back to Neuron X. X very quickly meets its threshold and fires to Neuron Y. 

A continual loop of activation that involved other neurons but the two or three neurons that continually stimulated each other would also fire to other neurons but those connections were not strong enough to cause the addition neurons to fire. Once another neuron would fire to the looping dual/triad of neurons, it could increase the activations strong enough to fire to other neurons outside of the looping neurons and thus potentially cause a chain of activations that would result in a behavior. 

I have yet to see any traditional AI with this level of recurrence and to be always on. It is typically only some sensory input that activates the AI to perform its analysis. I will discuss that we need to flip this idea of sensory input in a later post.

Continual stimulating neurons do cause issues with the emulations. Since I used UDP for neurons to talk to one another, eventually what I tagged as UDP Stacking would occur. What this means is that the number of messages being sent from one neuron to another would be happening so fast, that if another neuron sends a message to a highly stimulated neuron, it has to get in line and maybe far behind other messages already waiting to be received. I would see the individual neuron application emulations eventually start to become kind of erratic and it is because of this UPD Stacking that it happened. I'll discuss other emulation methods in a later post.

Therefore, I conclude that recurrence is key to AGI. You change "thoughts" many, many times in the course of a few seconds. Humans and higher-level animals have the capacity of focusing whereas a worm only reacts to its environment. Even if I am concentrating on writing this blog post, I have had a large number of other thoughts cross in my mind, everything from a video I need to watch to when is breakfast because I am feeling a little hungry. If I were a GOFAI, I would only be writing this blog and no other thoughts or ideas would surface.