SuperNets

Summary - Brains do not consist of a single neural network. Instead, they are composed of many neural networks that share the load in carrying out various cognitive tasks. Heretofore, neural network researchers have built so-called hierarchical cascade and deep learning architectures by manually connecting a few static neural networks to one another to solve moderately ambitious classification problems. Now IEI has achieved a new kind of self-assembling neural cascade called a "SuperNet," in which individual neural networks, in the form of STANNO modules, autonomously interconnect to serve as both classifiers and idea-generators. In distinct contrast to traditional cascades, the individual neural networks train in situ, adapting to newly arriving data from the environment and learning from one another. ...In short, this is how IEI can and has quickly built synthetic brains capable of human level discovery and invention.

Details - Typically, when a neural network aficionado talks about neural networks, they are speaking of either individual neural networks, or so-called cascade structures in which a series of smaller neural networks have been connected to one another by a computer programmer or hardware designer. In contrast, IEI provides a very novel form of artificial neural network cascade in which the individual, STANNO-based neural networks and Creativity Machines autonomously connect themselves with one another to form very complex and contemplative neural architectures. Whereas individual neural networks may absorb complex memories and relationships, these "networks of neural networks," that we call SuperNets, can absorb even more ambitious memories and relationships. Further, they can originate new ideas and plans of action, because they incorporate Creativity Machine based modules. Because all the component neural networks within the SuperNet are based upon STANNO technology, they continue to learn in situ, without recourse to a centralized training algorithm. For all intents and purposes, the SuperNet may be thought of as a neural network whose neurons are whole neural nets, or Creativity Machines that are constantly broadcasting ideas and plans of action to other neural networks within the cascade.

SuperNets are a significant tool in IEI's patented neural network toolbox, allowing us to 'grow' synthetic cognitive structures capable of perception, learning, and creativity. They may even equip themselves with the attitude (i.e., the self-perception) that they serve some significant purpose. So, whereas Madison Avenue types have stretched the truth before about software or hardware that 'thinks', the IEI SuperNets come the closest of any preceding AI technology to true cognition and consciousness. Further, when they are dissected, we find that all cognitive tasks have been loosely delegated among certain neural networks within the architecture, just as in the brain.

The principal use of SuperNets has been in the creation of highly advanced machine vision systems in which STANNOs autonomously dock and connect with one another to build the equivalent of the visual pathways of the brain. In this way, raw pixel patterns output by state of the art cameras may be rapidly processed to classify objects and scenarios within the camera's field of view, or detect anomalies even within dynamically changing scenes. Another revolutionary use of SuperNets have been in the self-organization of synthetic brains for robots. Effectively SuperNet based control systems can originate from neural nets that have totally no experience or learning. Once given the go-ahead, these creative robotic brains self-organize to achieve both near- and far-term strategies for fulfilling loosely defined objectives.

Finally, SuperNets are the forerunners of DABUS, wherein concepts are not expressed as neuron activation patterns, but the geometries autonomously forming neural modules.