Susan A. Werness

1994 Special Issue: Biological plausibility of synaptic associative memory models

Observations in brains of neuronal networks that subserve associative learning in living organisms have been exceedingly sparse until the past decade. Recently, some fundamental biophysical and biochemical properties of biological neural networks that demonstrate associative learning have been revealed in the marine mollusc. Hermissenda crassicornis. In mammals, we have localized distributed changes, specific to associative memory, in dendritic regions within biological neural networks. Based on these findings, it has been possible to construct an artificial neural network, Dystal (dynamically stable associative learning) that utilizes non-Hebbian learning rules and displays a number of useful properties, including self-organization; monotonic convergence; large storage capacity without saturation; computational complexity of O(N); the ability to learn, store, and recall associations among arbitrary, noisy patterns after four to eight training epochs; a weak dependence on global parameters; and the ability to intermix training and testing as new training information becomes available. The performance of the Dystal network is demonstrated on problems that include face recognition and hand-printed Kanji classification. The computational linearity of Dystal is demonstrated by its performance on a MasPar parallel hardware computer.

Associative learning in a network model of Hermissenda crassicornis

A companion paper in a previous issue of this journal presented a resistance-capacitance circuit computer model of the four-neuron visual-vestibular network of the invertebrate marine mollusk Hermissenda crassicornis. In the present paper, we demonstrate that changes in the models output in response to simulated associative training is quantitatively similar to behavioral and electrophysiological changes in response to associative training of Hermissenda crassicornis. Specifically, the model demonstrates many characteristics of conditioning: sensitivity to stimulus contingency, stimulus specificity, extinction, and savings. The models learning features also are shown to be devoid of non-associative components. Thus, this computational model is an excellent tool for examining the information flow and dynamics of biological associative learning and for uncovering insights concerning associative learning, memory, and recall that can be applied to the development of artificial neural networks.

Parametric analysis of dynamic postural responses

A detailed theoretical understanding of postural control mechanisms must be preceded by careful quantification of both the deterministic and stochastic aspects of postural behavior of normal and abnormal subjects under various dynamic conditions. Toward this end, concise parametric transfer function plus noise models were derived for both shoulder and waist position data obtained by applying a linear anteriorposterior bandlimited pseudorandom disturbance to the base of support of human subjects. Model orders as well as model parameters were determined empirically. One advantage of this modeling procedure is the conciseness of the postural models, permitting easy statistical analysis of the data obtained under different dynamic conditions from many subjects. Model features, including pole and zero locations, from 6 normal subjects each tested on 5 consecutive days under 3 input amplitudes and eyes open and closed conditions are presented. The resulting transfer function models consist of only 1 or 2 poles near the integration position on the Z plane unit circle and 0 to 2 zeros. Locations of the poles indicate that the eyes closed responses are more oscillatory, less damped, and with higher gains than the eyes open responses. These transfer functions are similar to nonparametric ones of other authors. The noise model orders are also small. Their spectra are those of low pass systems. Also, the quantity and frequency range of the postural noise is positively related to the amplitude of platform motion as well as related to the presence or absence of vision.A detailed theoretical understanding of postural control mechanisms must be preceded by careful quantification of both the deterministic and stochastic aspects of postural behavior of normal and abnormal subjects under various dynamic conditions. Toward this end, concise parametric transfer function plus noise models were derived for both shoulder and waist position data obtained by applying a linear anteriorposterior bandlimited pseudorandom disturbance to the base of support of human subjects. Model orders as well as model parameters were determined empirically. One advantage of this modeling procedure is the conciseness of the postural models, permitting easy statistical analysis of the data obtained under different dynamic conditions from many subjects. Model features, including pole and zero locations, from 6 normal subjects each tested on 5 consecutive days under 3 input amplitudes and eyes open and closed conditions are presented. The resulting transfer function models consist of only 1 or 2 poles near the integration position on the Z plane unit circle and 0 to 2 zeros. Locations of the poles indicate that the eyes closed responses are more oscillatory, less damped, and with higher gains than the eyes open responses. These transfer functions are similar to nonparametric ones of other authors. The noise model orders are also small. Their spectra are those of low pass systems. Also, the quantity and frequency range of the postural noise is positively related to the amplitude of platform motion as well as related to the presence or absence of vision.

Automated two- and three-dimensional, fine-resolution radar imaging of rigid targets with arbitrary unknown motion

An automated system for the SAR/ISAR imaging of rigid bodies which are undergoing arbitrarily complicated unknown motions is being developed. This system determines, from only the radar data, all observable parameters of motion, on a pulse by pulse basis. The approach makes it possible to: (1) exploit any type of relative motion: translational, rotational, two dimensional, three dimensional, deterministic, or stochastic; no prior parametric assumptions on the functional form of the motion are required; (2) require only the radar data; no ancillary motion measurement system on either the radar platform or on the target is required; (3) automatically provide all the motion information needed to form correctly scaled images, without cross range scale ambiguities; (4) make full use of all the radar data; no signals returning from a target are discarded; and (5) require a known computation time, which is not signal dependent, as all iterative processes used have known, guaranteed convergence rates.

Radio frequency interference removal in a VHF/UHF deramp SAR

This paper summarizes the results of an ARPA/Army sponsored program to develop innovative approaches for reducing the effects of multiple source of radio frequency interference (RFI) on synthetic aperture radars (SARs) operating in the frequency range of 100 MHz to 1000 MHz. Since the SAR signal can be modeled as wide band noise, the approach taken to achieve the objective was to model the RFI as a collection of tones within the desired SAR signal bandwidth. RFI suppression consisted of detecting the presence of and estimating the number of these tones, estimating their amplitude, phase, and frequency, and finally reconstructing the RFI and coherently subracting it from the original corrupted signal. Central to our approach was the use of a parametric maximum likelihood (PML) algorithm for the estimation of the parameters of the RFI tones. Although most of our effort was devoted to the evaluation of the performance obtainable from the PML algorithm, a variation of band-stop filtering, which is referred to as the Notch or Adaptive Mask algorithm, was also studied. Since the focus of this program was the development of algorithms for the ultra-wideband (UWB) P-3 SAR, which is a deramp SAR, a means of applying the PML algorithm to deramped RFI was also necessary. This paper will thus briefly describe the PML algorithm and how it can be applied to a deramp SAR, and it will then discuss the preformance of both RFI suppression algorithms and their computational complexity. As a result of this one year effort, two RFI algorithms have been developed that automatically remove 90 to 95 percent of the RFI that could have corrupted a SAR image.

Statistical evaluation of predictive data compression systems

A simple image reconstruction evaluation procedure has been developed for use in analysis and design of image compression systems. The evaluation consists of two parts: 1) examination of the autocorrelation function of the reconstruction errors, and 2) comparison of the distribution size and shape of the reconstructed image to that of the original. The philosophy behind the evaluation procedure is rooted in consideration of visual mechanisms and in linear system identification model validation techniques. Although originally postulated for use in the development of compression systems for noisy synthetic aperture radar (SAR) imagery for which the usual mean square error criterion is particularly useless, the evaluation procedure is proposed to be useful for analysis of any image compression system. The utility of the procedure is demonstrated with the selection of the best quantizer step sizes and data rates for an SAR predictive coding algorithm combined with a switched quantizer. It is also demonstrated with SAR data from which the speckle noise has been removed.

Computational Hermissenda Photoarray Model

Some of the simpler invertebrates demonstrate robust and efficient features of sensory information processing, integration, and motor control. Development and analysis of computer models of the essential neural mechanisms leading to these capabilities elucidate design principles for flexible man-made remote sensing, vision, and control systems. The marine snail, Hermissenda crassicornis, is ideal for study and modeling of biological computing mechanisms because the connectivity of its visual and vestibular neuronal network has been determined in considerable detail. To extract principles of sensory integration, a computer model of the photoarray in a single Hermissenda eye is developed and analyzed. This model demonstrates 1) information processing capabilities observed in Hermissenda’s visual sensors, and 2) the importance of heterogeneity in the efficient and robust operation of biological networks.

Application Of Predictive Compression Methods To Synthetic Aperture Radar (SAR) Imagery

Several variations of a prediction compression system have been demonstrated with 6 meter resolution synthetic aperture radar (SAR) imagery. Due to the uncorrelated nature of SAR imagery, the prediction system design problem was approached from the point of view of statistics matching and decorrelation of reconstruction errors, rather than minimization of mean square error. It is shown that a moving average (MA) predictor can work well, depending upon the quantizer ,used and upon the homogeneity of the data. Due to the occurrence of large data values evolving from returns from cultural objects, slope overload can be a severe problem in system design. This problem is most economically solved by a thresholding type of operation in the quantizer, resulting in a dual rate system. Good results are obtainable at rates of 1.5 bits/pixel.