Abduction, inference to the best explanation, is an information-processing task that is useful for solving interpretation problems such as diagnosis, medical test analysis, legal reasoning, theory evaluation, and perception. The task is a generative one in which an explanation comprising of domain hypotheses is assembled and used to account for given findings. The explanation is taken to be an interpretation as to why the findings have arisen within the given situation. Research in abduction has led to the development of a general-purpose computational strategy which has been demonstrated on all of the above types of problems. This abduction strategy can be performed in layers so that different types of knowledge can come together in deriving an explanation at different levels of description. Further, the abduction strategy is tractable and offers a very useful tradeoff between confidence in the explanation and completeness of the explanation. This paper will describe this computational strategy for abduction and demonstrate its usefulness towards perceptual problems by examining problem-solving systems in speech recognition and natural language understanding.
Collective self-organization effects and chaos are commonly observed in optics. We describe examples in a particular kind of nonlinear optical material: photorefractive crystals. In particular, we show different effects that arise when photorefractive crystals are illuminated by one laser beam, two laser beams, and three laser beams.
It is well known that patient motion causes artifacts that can mimic disease and lead to mis-diagnosis. Various physiological gating methods have been investigated in the past to combat this problem by acquiring CT scans during the quiescent motion periods. Previously, we proposed a predictive gating algorithm for computing the quiescent time intervals to automatically starting the scanner. The algorithm uses adaptive moving correlation to exploit the fact that the shape of the inspiratory or expiratory segment of the waveform is similar from breath to breath. The CT data acquisition is triggered when the correlation coefficient exceeds a predefined threshold. Although this method performs satisfactorily in most patients, it fails to trigger CT scans in some patients when excessive variation in the motion waveform exists. To overcome this difficulty, we propose an improved algorithm that will determine the correlation threshold based on the waveform history acquired during the patient preparation. We further exclude the portions of the breathing curve that deviate significantly from the average breathing curve based on the low correlation coefficients. We then calculate the weighted correlation coefficients with the most recent samples carrying higher weights. The start of a CT scan is then determined based on the weighted average. Various experiments have demonstrated the advantages and effectiveness of our approach.
Manufacturing is usually defined as the process of turning raw materials into useful things. It is a set of actions performed on the materials. The materials are passive. Manufacturing is done to the materials. We explore here manufacturing in which the material itself plays an active role.
2D quasi-crystals were fabricated from polystyrene microspheres and characterized for their structural, diffraction, and non-linear optics properties. The quasi- crystals were produced with the method based on Langmuir- Blodgett thin film technique. Illuminating the crystal with the laser beam, we observed the diffraction pattern in the direction of the beam propagation and in the direction of the back scattering, similar to the x-ray Laue pattern observed in regular crystals with hexagonal structure. The absorption spectrum of the quasi-crystal demonstrated two series of regular maxima and minima, with the spacing inversely proportional to the microspheres diameter. Illumination of the dye-doped microspheres crystal with Q- switched radiation of Nd:YAG laser showed the enhancement of non-linear properties, in particular, second harmonic generation.
We show that one can treat pseudo-random generators, evolutionary models of texture images, iterative local adaptive filters for image restoration and enhancement and growth models in biology and material sciences in a unified way as special cases of dynamic systems with a nonlinear feedback.
An adaptive method for small-size image recognition is suggested; it is based on iterative procedure of determining the parameters of a feature space. A structure of optical- electronic systems with neural network elements is considered. A high discrimination properties of the formed features stipulated by using both the values of samples and the information on the structure of auto- and cross- correlation functions in the iterative process was verified as a result of numerical simulation. Invariance of the method to various illumination conditions and to some extent to aspect transformation s of recognized images is shown.
The biological neuronal model is a n electrically polarized membrane capacitor with numerous ionic channels for synaptic inputs and leakage currents. It is generally a complex system whose geometrical, electronic, and chemical substructures synergistically determine its temporal dynamics in response to its input distributions. The integrate-and-fire pulse generator resets a local internal voltage activity, which in turn interacts with the other voltage distributions in the dendrites. A recent closed-form solution to a compartmental neuronal model shows that it is a fully chaotic system with algorithmic properties that provide a new approach to object-oriented data processing.
Object isolation in an image is an iterative process in which the target gradually fade. The intent of this design is to be more robust to noise, object alterations, and clutter than traditional filters. This iterative process can be performed by a non-linear shunting pulse-coupled neural network (PCNN). The input scene is modified by the filtered pulse response of the PCNN. Within several iterations the target is enhanced and the non-targets are degraded to such a degree that the target is completely dominant in the input. The segmented pulse response of the PCNN allows for a unique approach to this system.
The pulse-coupled neural network (PCNN) is a biologically motivated algorithm that has tremendous potential. The PCNN has several inherent properties that make it useful in a variety of applications. This paper introduces a syntactical information processing architecture that is mainly driven by PCNNs. It exploits several of the inherent properties of the PCNN to receive and process information, to convert the information to a syntactical phrase and its 1D form which can then be used to generate an appropriate response to the information received as an input. Examples of PCNN performance which is appropriate for this systems are given.
In the many years that pattern recognition has been of interest, there have ben many clever advances. One recent advance is the pulse-coupled neural network (PCNN). Due to recent developments in PCNNs, it is becoming increasingly possible to recognize images in space regardless of scale, rotation, translation. A continuation of this has been investigated which will allow images to be recognized audibly. In this paper, a general method for converting 2D spatial patterns into pattern-specific sound patterns will be discussed, along with some background information on PCNNs, and projections for his image-to-sound conversion.
Classical mechanics assumes that its laws are independent of spatio-temporal resolutions. To see whether there is an alternative to this assumption we write the energy of a relativistic particle in a finite-difference form.
In human eyes, retina cells have shown complicated, sophisticated structures, and thus capability of carrying out certain basic visual functions. In computer vision, a spherical model has been proposed to demonstrate that certain basic visual functions can be performed at the sensor/photoreceptor level for target recognition, mobile robotic control, and other applications. This paper will establish that low level computer vision in this spherical model exhibits a Hamiltonian model of quantum computation. This results will be important to the understanding of human vision as well as the development of intelligent visual sensor chips and devices in electronics and optics.
Recently, quantum computers have been proposed as potentially the next-generation computer systems which will overcome the physical and computational limitations of current solid state technology. This paper suggests that the natural logic of quantum computers should be the probabilistic logic rather than the ordinary logic. In probabilistic logic, computational problems formulated directly may be solved with much less complexities by quantum computers.
The theory and equivalental models of neural networks based on equivalence operation of continuous and multivalued neural logic are considered. Their connection with metric of metric-address spaces are shown. Normalized equivalences of vectors with multilevel components are determined. Equivalental models for simple network with weighted correlation coefficients, for network with adapted weighing and double weighing are suggested. It is shown, that the network model with double weighing being most generalized can also conduct the recalculation process of networks to two-step algorithms without calculation of connections matrix. Equivalent models require calculations based on vector-matrix procedures with equivalence operation and can be realized on vector-matrix calculations based on vector- matrix procedures with equivalence operation and can be realized on vector-matrix equivalentors with space and time integration. The apparatus implementations of models with productivity of 108 divided by 109 connections/sec and neuron number 256 and more are suggested.
The need for autonomous systems to work under unanticipated conditions requires the use of smart sensors. High resolution systems develop tremendous computational loads. Inspiration from animal vision systems can guide us in developing preprocessing approaches implementable in real time with high resolution and deduced computational load. Given a high quality optical path and a 2D array of photodetectors, the resolution of a digital image is determined by the density of photodetectors sampling the image. In order to reconstruct an image, resolution is limited by the distance between adjacent detectors. However, animal eyes resolve images 10-100 times better than either the acceptance angle of a single photodetector or the center-to-center distance between neighboring photodetectors. A new model of the fly's visual system emulates this improved performance, offering a different approach to subpixel resolution. That an animal without a cortex is capable of this performance suggests that high level computation is not involved. The model takes advantage of a photoreceptor cell's internal structure for capturing light. This organelle is a waveguide. Neurocircuitry exploits the waveguide's optical nonlinearities, namely in the shoulder region of its gaussian sensitivity-profile, to extract high resolution information from the visual scene. The receptive fields of optically disparate inputs overlap in space. Photoreceptor input is continuous rather than discretely sampled. The output of the integrating module is a signal proportional to the position of the target within the detector array. For tracking a point source, resolution is 10 times better than the detector spacing. For locating absolute position and orientation of an edge, the model performs similarly. Analog processing is used throughout. Each element is an independent processor of local luminance. Information processing is in real time with continuous update. This processing principle will be reproduced in an analog integrated circuit using photodiodes and fiber optic waveguides as the nonlinear light sensing devices, current mirrors and opamp circuits for the processing. The outputs of this circuit will go to other artificial neural networks for further processing.
It was shown earlier that models of cortical neurons can, under certain conditions of coherence in their input, behave as recursive processing elements (PEs) that are characterized by an iterative map on the phase interval and by bifurcation diagrams that demonstrate the complex encoding cortical neurons might be able to perform on their input. Here we present results of numerical experiments carried on a recurrent network of such recursive PEs modeled by the logistic map. Network behavior is studied under a novel scheme for generating complex spatio-temporal input patterns that could range from being coherent to partially coherent to being completely incoherent. A nontraditional nonlinear coupling scheme between neurons is employed to incorporate recent findings in brain science, namely that neurons use more than one kind of neurotransmitter in their chemical signaling. It is shown that such network shave the capacity to 'self-anneal' or collapse into period-m attractors that are uniquely related to the stimulus pattern following a transient 'chaotic' period during which the network searches it state-space for the associated dynamic attractor. The network accepts naturally both dynamical or stationary input patterns. Moreover we find that the use of quantized coupling strengths, introduced to reflect recent molecular biology and neurophysiological reports on synapse dynamics, endows the network with clustering ability wherein, depending ont eh stimulus pattern, PEs in the network with clustering ability wherein, depending on the stimulus pattern, PEs in the network divide into phase- locked groups with the PEs in each group being synchronized in period-m orbits. The value of m is found to be the same for all clusters and the number of clusters gives the dimension of the periodic attractor. The implications of these findings for higher-level processing such as feature- binding and for the development of novel learning algorithms are briefly discussed.
In this paper, we studied the trajectory generation problem for a two-degrees-of-freedom robot in a workspace with obstacles. To generate the robot's trajectories, we developed a genetic algorithm to search for valid solutions in the configuration space. Our results present a novel perspective on the problem not seen in the conventional robot trajectory planners. The genetic algorithm approach is beneficial because it may be extended to plan trajectories for robots with more degrees of freedom. The evolutionary search process may allow the user to solve the trajectory problem in an n-dimensional space where the 'curse of dimensionality' inevitably stalls conventional methods. We demonstrate the algorithm with some examples and discuss the possible extension to higher order problems.
Due to the development of digital information system in medical field, a large amount of image or signal data obtained from health examination has been stored. Analyzing these data is expected to make it possible to formulate new diagnostic knowledge for health care. In this paper, we propose a classification method suitable for the analysis of a large amount of medical data, for the purpose of assisting medical doctors to analyze the data. Int he proposed method, image or signal data are treated as vectors and mapped into multi-dimensional space, then hierarchical clustering method is applied. To obtain optimal division of cluster, a statistical criterion is introduced, and a binary tree of clusters is constructed base don the criterion. From the results of experiment using generated data and ECG signal, it is confirmed that the data sets can be correctly classified by our proposed method.
We report on the application of genetic programming to the determination of a desired output vector from an input vector. Genetic programming is an emerging technique similar in spirit to genetic algorithms which employ a metric to drive a parallel search of the solution space. In contrast to genetic algorithms which yield a single encoded string as the solution, genetic programming yields a computer program which can be examined and understood. Genetic programming also offers the possibility of enabling a technique whereby feature vectors can be automatically developed. We have applied the technique to the determination of ship aimpoints from segmented imagery using input from a sensor. The raw imagery is then processed and a feature vector extracted, as was done in a previous problem. The feature vectors are then used as input to the genetic programming technique. We will report on the sensitivity of performance of the genetic programming technique as a function of the metric employed. In addition we will compare the performance of the computer program obtained by genetic programming to the performance of a back propagation neural networks developed for our problem. Furthermore we will report on the performance results obtained using genetic programming with and without the presence of automatically created subroutines as well as the determination of critical inputs.
In this paper, we will propose a method to represent multidimensional fuzzy terms on a particular hardware device, the processor W.A.R.P. 1 (weight associative rule processor). The choice of this hardware depends on its characteristics here described together with its structure. In order to explain the method and its advantages with respect to the traditional approach using aggregation operators, we will show how a simple fuzzy rules system, concerning a real robotics problem, is implemented in both cases.
This paper shows that the Vapnik-Chervonenkis (VC) dimension of a set of functions representing a single hidden-layer, feed-forward, single binary output processor artificial neural network (ANN) can be evaluated using the characteristic (Poincare) polynomial of the implied hyperplane arrangement. This is a significant result since it lays a mathematical foundation rooted in combinatorial geometry for measuring ANN capabilities. The ANN specified above, geometrically, is a hyperplane arrangement configured to dichotomize a signed set. Since it is known that he cut- intersection of the hyperplane arrangement is a semi- lattice, then the Poincare polynomial can be used to evaluate certain geometric invariants of this semi-lattice. One of these geometric invariants is the cardinality of the resultant chamber set of the arrangements, which this paper will show is the VC dimension. Given this connection, ANN capabilities can be characterized in more general terms of geometric invariants about the hyperplane arrangements and signed set configurations. In the case of the VC dimension, the invariant about the data is simply the cardinality of the set independent of the coloring or the geometric arrangement. Hence, VC dimension assumes a worst-case data configuration even though the requirements of an ANN architecture could vary dependent on the coloring and arrangement. With this relationship established between ANNs and combinatorial geometry, alternative geometric invariants can be investigated pacing the way for improving the capabilities and designs of ANN architectures for mathematical and physical systems.