I've noticed that most ANN's do not incoporate a lot of realistic variables from biological neurons. However, I have found that doing so allows for networks that are capable of much higher processing than current ANN's. For instance, a single biological neuron can add, or subtract an infinite number of bits together without error by inputting the bits in a continuous fashion (meaning temporally). A single neuron can also multiply an infinite string of bits to a fixed length of bits. These neurons are specifically designed to perform these operations, but they only use biological components. It doesn't use sigmoid output or integrate-and-fire. It does use the neuron threshold and spike output. It also uses dendritic synapses, dendritic thresholds, synaptic weights, and the STDP learning rule that has been discovered in neurons. It also uses something I've never seen in any other ANN: dendrite outputs (or dendrites that synapse on other dendrites).

Another thing I've found about realistic neurons is that memory capacity and memorization is a lot simpler. Real neurons have axonal conduction delays, thereby creating polychronization. Neurons also store information in an abstract form (meaning representing an entire memory with one neuron). Using both principles, I've found a theoretical and very promising method to high capacity, multi-associative (meaning 2 or more things associated together) memory. I've already gotten a single neuron to represent a memorized list of input signals (inputs sent over time). However, it has a fixed limit to the length it can contain. Hopefully, my new findings on polychronization will allow me to associate and store these abstract memories in a network that has more memory "slots" than synapses.

Now I have to wonder, why aren't the more biologically realistic models getting used more when the other ANN's have encountered their obvious limitations that even simple animal brains can exceed? Granted, the non-realistic ANN's do have their uses, I still believe that realistic ANN's are the best approach to AI in general. Not to mention, our brains and animal brains can already do things that we're trying to get non-realistic ANN's to do (such as object extraction from images, pattern recognition, high capacity, bi-directional associative memory, etc.). If we want powerful AI, I think we need to use more realistic ANN's. The real neural networks work.

The Intellector