HOME OF THE CREATIVITY MACHINE
The Big Bang of Machine Intelligence!
The IEI Blog is now open for discussion of new business ventures and opportunities!
IEI's Patented Group Membership Filters
Summary - IEI holds the patent on the use of a special kind of neural network called an auto-associative net to rapidly identify patterns that are similar to other patterns the network has previously been exposed to. Conversely, this methodology may be used to identify patterns that are non-representative of those already shown to the network. In either usage, we need only show the auto-associative network representative patterns of what we would like it to detect, without having to expose it to counterexamples. This patented technique has proven extremely effective in IEI's automated target recognition (ATR), machine vision, and semantic parsing programs. This methodology has also been applied to the company's network intrusion detection and fault detection systems used in both control and robotic applications developed with the Air Force Research Laboratory.
Details - IEI's patented group membership filter (GMF) is based upon a particular way to use auto-associative networks and are an outgrowth of research conducted by our founder in 1975. Such auto-associative networks have input and output layers having the same numbers of neurons. They learn through cumulative training to reconstruct input patterns at their output layers.
If such networks are trained upon patterns corresponding to a particular genre, such as the bitmaps of human faces, two important things happen: (1) The weights leading to the hidden layer(s), shown in blue in the diagram to the right, self-organize so as to detect the critical features of that genre, such eyes, ears, chins, and noses, and (2) The output layer weights, shown in red, self-organize so as to capture the inherent constraint relations that define the genre, such as eyes flank the nose, just above it, and the mouth is under the nose, etc.
If such a network is trained upon many bitmaps of faces and then shown a new facial bitmap, the appropriate features are detected at the hidden layer(s) and the necessary constraint relations, here topological in nature, are obeyed. The result is that the input pattern, P, is faithfully reconstructed at the output layer, as the output pattern P'. If, on the other hand, the bitmap of a foot, for example, is applied as pattern P, the critical features are not recognized, nor are the topological constraints, and the input pattern is not successfully reconstructed. Whether the input pattern represents the genre or group G, absorbed through training into the network, the difference vector, P-P', indicates how representative the input pattern is to what the network has previously learned in training.
Note that in contrast to hetero-associative networks, no counterexamples are needed for training. The network simply learns to recognize similar things to which it has been previously exposed.
Similarly, the GMF may be used as an anomaly detector. For instance, it may be trained upon the normal behavior of some system, such as signals from all the sensors in a chemical reactor during its routine operation. As the GMF watches and learns the fundamental behavior of the system, it forms a comprehensive model of it. As any abnormalities arise in the system being monitored, the difference vector, P-P', rises. The most anomalous of components in this vector quickly indicate which sensors are involved in the anomaly. Thereafter, rule-based systems, or even better, Creativity Machines, may now move the system toward a state that assures the desired process outcome.
If the patented GMF / anomaly detection scheme is now implemented via the likewise patented Self-Training Artificial Neural Network Object or STANNO, we attain classification and anomaly detection schemes that may train and classify patterns having millions of inputs and outputs, on ordinary Pentium class processors, on timescales of the order of milliseconds. Furthermore, GMFs may autonomously connect themselves into brain-like structures, or SuperNets, that perform even more complex classification and anomaly detection tasks.
Currently, STANNO-based GMFs and anomaly detectors form the basis of our network intrusion and fraud detection systems, advanced machine vision applications, and creative robots.
Thaler, S. L. (2000). US Patent 6,014,653, Non-Algorithmically implemented artificial neural networks and components thereof.
Thaler, S. L. and Conrad,
(1998). Real-Time Fault Detection Using
Auto-associative Filtering, AIRTC, Oct. '98.
© 1997-2016, Imagination Engines, Inc. | Creativity Machine®, Imagination Engines®, Imagitron®, and DataBots® are registered trademarks of Imagination Engines, Inc.
1550 Wall Street, Ste. 300, St. Charles, MO 63303 • (636) 724-9000