Browsing by Author "Teolis, A."
Results Per Page
Sort Options
Item Adaptive Pattern Classification(1989) Teolis, A.; Shamma, S.; ISRUp until the recent past, the power of multi layer feed forward artificial neural networks has been untapped mainly due to the lack of algorithms to train them. With the emergence of the backpropagation algorithm; however, this deficiency has been removed. Despite this innovation, the backpropagation method is still not without its drawbacks. Among these the most prominent are the facts that i) the learning is conducted in a supervised manner and ii) that learning and operation must occur in two distinct phases. Because of these properties, the backpropagation algorithm falls short of solving a 'true' pattern classification problem. This is not to say that a network could not be trained via backpropagation to mimic a previously solved pattern classification scheme; but that the backpropagation method is not capable of autonomously generating classification schemes. A more realistic (and certainly more useful) learning scenario is that patterns would be presented without supervision to the system continuously; consequently, the system will begin to group like patterns into similar classes and continue to do so indefinitely; i.e. continuous learning. It is exactly this type of learning that is discussed here.Item Cascaded Neural-Analog Networks for Real Time Decomposition of Superposed Radar Signals in the Presence of Noise.(1989) Teolis, A.; Pati, Y.C.; Peckerar, M.C.; Shamma, S.; ISRAmong the numerous problems which arise in the context of radar signal processing is the problem of extraction of information from a noise corrupted signal. In this application the signal is assumed to be the superposition of outputs from multiple radar emitters. Associated with the output of each emitter is a unique set of parameters which are in general unknown. Significant parameters associated with each emitter are (i) the pulse repetition frequencies, (ii) the pulse durations (widths) associated with pulse trains and (iii) the pulse amplitudes: A superposition of the outputs of multiple emitters together with additive noise is observed at the receiver. In this study we consider the problem of decomposing such a noise corrupted linear combination of emitter outputs into an underlying set of basis signals while also identifying the parameters associated with each of the emitters involved. Foremost among our objectives is to design a system capable of performing this decomposition/classification in a demanding realtime environment. We present here a system composed of three cascaded neural-analog networks which, in simulation, has demonstrated an ability to nominally perform the task of decomposition and classification of superposed radar signals under extremely high noise conditions.Item Classification of the Transient Signals via Auditory Representations(1991) Teolis, A.; Shamma, S.; ISRWe use a model of processing in the human auditory system to develop robust representations of signals. These reduced representations are then presented to a neural network for training and classification.Empirical studies demonstrate that auditory representations compare favorably to direct frequency (magnitude spectrum) representations with respect to classification performance (i.e. probabilities of detection and false alarm). For this comparison the Receiver Operating Characteristic (ROC) curves are generated from signals derived from the standard transient data set (STDS) distributed by DARPA/ONR.
Item Discrete Representation of Signals from Infinite Dimensional Hilbert Spaces with Applications to Noise Suppression and Compression(1993) Teolis, A.; Benedetto, J.J.; Shamma, S.A.; ISRAddressed in this thesis is the issue of representing signals from infinite dimensional Hilbert spaces in a discrete form. The discrete representations which are studied come from the irregular samples of a signal dependent transform called the group representation transform, e.g., the wavelet and Gabor transform. The main issues dealt with are (i) the recoverability of a signal from its discrete representation, (ii) the suppression of noise in a corrupted signal, and (iii) compression through efficient discrete representation.The starting point of the analysis lies with the intimate connection between the Duffin-Schaeffer theory of (global) frames and irregular sampling theory. This connection has lead elsewhere to the formulation of iterative schemes for the reconstruction of a signal from its irregular samples. However, these schemes have not addressed such issues as digital implementability and reconstruction from perturbed representations. Here, iterative reconstruction algorithms are developed and implemented which recover a signal from its possibly perturbed discrete representation.
Robustness to perturbations occurring directly in the signal domain are also investigated. Based on the notion of coherence with respect to a frame, a simple non-linear thresholding scheme is developed for the rejection of noise.
The structure of the discretization has many free parameters including the choice of group representation transform, the analyzing function associated with the group representation transform, and the sampling set. Each choice of parameters leads to a different discrete representation and the specification of an underlying set of primitive functions. Reconstructability is directly related to the frame properties of this set of primitive functions.
Localized discrete representations around a particular signal are also investigated. Truncations and other signal dependent localization of global representations lead to finite representations. The approach to finite representations which is taken here can be stated in terms of local frames for the reproducing kernel Hilbert Space formed by the range of the group representation transform.
Finally, numerical examples of discrete representations which are signal independent and new signal dependent discrete (positive extreme) wavelet representations are presented. Reconstruction, noise suppression, and compression experiments are conducted and demonstrated on numerical examples including speech and synthetic signals.
Item A Graphical Simulation Management System(1991) Teolis, A.; ISRIn this paper we describe the design and implementation of a system for the management of complex interactive graphical dynamical simulations. The main objective of the system is to provide tools which enable a graphical simulation to be built with minimum effort. These tools include routines for building interfaces (panels consisting of buttons, sliders, type in boxes, etc.), arranging the simulation workspace, describing complex 3D objects in terms of predefined (or user defined) primitives, and providing objects with properties used in conjunction with lighting models.The need for such a management system is supported by the facts that both the development effort and the extent of machine-specific knowledge needed to successfully implement an arbitrary graphical simulation is significant. It is the aim of the Graphical Simulation Management System (GSMS) to reduce these burdens and place the power of 3D simulation within easy reach.
We have presented several examples of dynamical simulations in which the power and diversity of the management system is demonstrated.