Parametric Avalanche Stochastic Filter
Temporal pattern recognition is the recognition of patterns that may or may not be stable over time. Why is this important? In real life, very little is stable and unchanging, particularly when the objective is to detect some event or action, process it with a computer, and act upon the event or action. Speech recognition, control systems (such as numerically-controlled machine tools), fault detection, and video processing to name a few, are all examples of dynamical systems. Dealing with these types of real-world systems requires special techniques for computer processing.
Martingale Research has developed a neurocomputing architecture called the Parametric Avalanche Stochastic Filter (PASF). This architecture can be used to provide robust solutions to a wide variety of complex problems not achievable by conventional computer architectures. The PASF architecture is based on the mathematical methods of Quantum Neurodynamics to solve several fundamental problems of dynamical and cognitive systems. Patent filings for the PASF have been executed in the U.S., Japan, Australia, and the EEC.Back to top
The PASF is based on Quantum Neurodynamics (QND), a mathematical theory developed specifically to deal with real-world phenomena. QND applies methods of quantum mechanics to nonlinear dynamical systems to create a scalable, computationally efficient algorithm for observation and control of nonlinear systems. By using well-understood equations originally developed to describe the evolution of subatomic physical systems, QND is able to efficiently model and process more general dynamical systems that would otherwise require far more computational work.
QND provides an adaptive modeling and stochastic filtering algorithm for Hamiltonian systems in non-Gaussian noise. It is inherently capable of dealing with non-stationary dynamic inputs, and is excellent at noise suppression and the detection of changes in an input sensor stream. The PASF is a realization of this concept, and illustrates how diverse its applications can be.Back to top
Parametric Avalanche Stochastic Filter
The PASF is a realization of QND in a neural network architecture. It excels in adaptive control environments and in processing real-time data, and has been applied to multiple types of dynamical systems.
- Adaptive stabilization of tethered satellite systems
- Automatic target detection and tracking of objects in dynamic imagery
- Tracking of multiple simultaneous targets
- Stabilization of dynamic nonlinear systems during disturbances
- Word recognition from continuous speech
- Processing of real-time image data
Beam Stabilization and Speech Recognition
Martingale Research has developed under U.S. Army sponsorship a control algorithm for dynamic stabilization of a flexible beam mounted on a moving platform with active sources of disturbance and nonlinearity. This project is also applying the PASF to large vocabulary continuous speech recognition in the presence of extreme noise. Subsequent applications of this technology include real-time adaptive controls, field-usable speech recognition systems, and intelligent front ends to existing systems.
Adaptive Controller for Tethered Satellites
Making use of the ability of the PASF architecture to stabilize complicated nonlinear systems, Martingale Research developed a prototype of a general purpose adaptive controller for NASA. The adaptive controller was specifically developed for the control and stabilization of tethered satellites, for which the only control input available is that of the tether. This is a very difficult problem for traditional approaches. However, instead of building a model of the dynamical system representing the satellite and processing large numbers of state variables, the PASF-based controller is able to use a single robust equation to calculate the evolution of the system. This algorithm is more efficient and more able to deal with complexities in the system.
This approach is widely applicable to problems requiring intelligent adaptive control, such as vehicle control systems and robotics. It also has applications anywhere a large system must be monitored and controlled, such as the automation of chemical and manufacturing processes.
Target Tracking System
Martingale Research has also applied the PASF to a target-tracking system developed for the Naval Surface Warfare Center. Designed to be inherently parallelizable and scalable, the system detects and tracks targets in dynamic imagery which may be from video, infrared, or radar sources. This system is capable of tracking multiple maneuvering targets accurately and in the presence of heavy noise. Martingale Research's approach provides the ability to begin tracking a target before sufficient likelihood is actually accumulated to declare that the target has been detected (track before detect), and can track multiple "probable" targets this way simultaneously. Applications of these results include intelligent sensors, security systems, real-time video processors, and solutions to other problems involving dynamic multidimensional data.Back to top Skip to navigation