Neural Learning System

Many of the traditional approaches to AI have failed because none of them were essentially different from the original computing paradigm: programming a very specific series of specific instructions for the computer to follow in precise manner and precise order. Genuine intelligence or even adaptive and robust behavior is unlikely to be achieved using this approach to problem solving. The cognitive functions of our brain must rely on tremendously more dynamic and self-architecting processes. During the course of his graduate and postdoctoral research Ayala studied the patterns and processes of molecular evolution, whereby the gradual accumulation of small modifications at the genetic level can produce highly complex and specialized adaptations at the organismal level. There are many exciting parallels between molecular evolution and neural learning: for example, between the dynamics of genetic alleles drifting in populations and the dynamics of competing neural firing patterns, between the emergence of complex genetic regulatory pathways and the emergence of complex interacting neural circuits, and between the fitness function of natural selection and the fitness function of positive and negative reinforcement and internal consistency.

This neural learning system is a significant departure from conventional artificial neural network models. It is controlled by over 50 separate parameters optimized to find the balance between the acquisition and retention of stored memories and the assimilation and processing of input data. Multifurcating and self-reinforcing feedback loops within the network are the functional repositories of memory. Metaloops -- coordinate interactions among these loops -- represent the integration of data and concepts within the network, and form stable peaks on an adaptive landscape. Competition among the metaloops and emergent neural circuits causes the network to make its own heuristic search through the universe of possible connection weights to find local optima. The network is trainable at any time through contingencies of positive and negative reinforcement. A Java program with dozens of classes and hundreds of pages of code implements this neural learning system, and a sample screen shot is shown on the left.

Several important emergent features of the neural system have been optimized from the development of a system of equations that characterize the network architecture. For example, the expected average length of feedback loops within the network is shown on the left as function of n, the number of nodes, and k, the number of connections per node. Other network features that have been optimized from these equations include the expected total number of feedback loops and the expected connection load, defined as the average number of loops in which each individual connection participates.

Additional network parameters, such as the neuronal threshold and the ratio of inhibitory to excitatory neurons, may be included in the optimization functions to generate dynamic attractors representing a stable preferred number of neuronal firings. In the graph on the left, the x-axis shows the number of excitatory firings and the y-axis shows the number of inhibitory firings. Each blue crosshair (+) represents the total firings within the network at any given instant, and is connected to a red asterix (*) showing the expected number of firings at the following instant. The yellow circle outlines the location of a stable attractor, illustrating that from any state of neuronal activity this network will soon settle on approximately six excitatory firings and one inhibitory firing, thus preventing the network from either running out of control from too much activity or dying out from too little. During training the number of firings may be nudged to a different location on the graph by manipulating certain network parameters, and while the amount of activity will return to that indicated by the location of the attractor, a different set of neurons will likely be firing, and thus a different set of coordinated metaloops will be activated.