Engineering topics
Publications and source records attributed to Toomarian, Nikzad.
Neural algorithm for simulation of class of nonlinear wave phenomena devised. Numerically solves special one-dimensional case of Korteweg-deVries equation. Intended to be executed rapidly by neural network implemented as charge-coupled-device/charge-injection device, very-large-scale integrated-circuit analog data processor of type described in "CCD/CID Processors Would Offer Greater Precision" (NPO-18972).
Charge-coupled-device/charge-injection-device (CCD/CID) data processors of proposed type offer advantages of massively parallel computational architecture and high computational speed typical of older CCD/CID data processors, but with increased precision. Useful in performing matrix vector multiplications in variety of applications, including solving partial differential equations, processing signal and image data, control computations, and neural-network simulations. Greater precision of proposed devices help to ensure accuracy in CCD/CID implementations of pseudospectral neural networks - particular class of artificial neural networks especially suited to solving nonlinear differential equations.
Improved method of adjoint-operator learning reduces amount of computation and associated computational memory needed to make electronic neural network learn temporally varying pattern (e.g., to recognize moving object in image) in real time. Method extension of method described in "Adjoint-Operator Learning for a Neural Network" (NPO-18352).
A neural network is trained to output a time dependent target vector defined over a predetermined time interval in response to a time dependent input vector defined over the same time interval by applying corresponding elements of the error vector, or difference between the target vector and the actual neuron output vector, to the inputs of corresponding output neurons of the network as corrective feedback. This feedback decreases the error and quickens the learning process, so that a much smaller number of training cycles are required to complete the learning process. A conventional gradient descent algorithm is employed to update the neural network parameters at the end of the predetermined time interval. The foregoing process is repeated in repetitive cycles until the actual output vector corresponds to the target vector. In the preferred embodiment, as the overall error of the neural network output decreasing during successive training cycles, the portion of the error fed back to the output neurons is decreased accordingly, allowing the network to learn with greater freedom from teacher forcing as the network parameters converge to their optimum values. The invention may also be used to train a neural network with stationary training and target vectors.
The emergence, in recent years, of hyperspectral sensors provides a tremendous opportunity for advancing the process of detailed and direct remote detection and identification from space of targets or surface materials. Such sensors exploit the uniqueness of the corresponding spectral reflectance signatures, which enables high resolution imaging spectrometer data to be processed on a pixel-by-pixel basis. This has implications both for defense-related applications (e.g., surveillance tasks) and in the civilian domain (e.g., for science applications). The purpose of this talk is to discuss a number of strong arguments that support neural networks as a choice for the generalized analysis (e.g., unmixing) of remotely sensed hyperspectral data.
Included in this viewgraph presentation on intelligent neuroprocessors for launch vehicle health management systems (HMS) are the following: where the flight failures have been in launch vehicles; cumulative delay time; breakdown of operations hours; failure of Mars Probe; vehicle health management (VHM) cost optimizing curve; target HMS-STS auxiliary power unit location; APU monitoring and diagnosis; and integration of neural networks and fuzzy logic.
This invention provides a new hierarchical approach for supervised neural learning of time dependent trajectories. The modular hierarchical methodology leads to architectures which are more structured than fully interconnected networks. The networks utilize a general feedforward flow of information and sparse recurrent connections to achieve dynamic effects. The advantages include the sparsity of units and connections, the modular organization. A further advantage is that the learning is much more circumscribed learning than in fully interconnected systems. The present invention is embodied by a neural network including a plurality of neural modules each having a pre-established performance capability wherein each neural module has an output outputting present results of the performance capability and an input for changing the present results of the performance capabilitiy. For pattern recognition applications, the performance capability may be an oscillation capability producing a repeating wave pattern as the present results. In the preferred embodiment, each of the plurality of neural modules includes a pre-established capability portion and a performance adjustment portion connected to control the pre-established capability portion.
This paper presents the implementation of a new class of massively parallel algorithms for solving certain time-dependent partial differential equations (PDEs) on massively parallel-supercomputers.
Electronic neural networks made to synthesize initially unknown mathematical models of time-dependent phenomena or to learn temporally evolving patterns by use of algorithms based on adjoint operators. Algorithms less complicated, involve less computation and solve learning equations forward in time possibly simultaneously with equations of evolution of neural network, thereby both increasing computational efficiency and making real-time applications possible.
An innovative hybrid, analog-digital charge-domain technology, for the massively parallel VLSI implementation of certain large scale matrix-vector operations, has recently been introduced. It employs arrays of Charge Coupled/Charge Injection Device cells holding an analog matrix of charge, which process digital vectors in parallel by means of binary, non-destructive charge transfer operations. The impact of this technology on massively parallel processing is discussed. Fundamentally new classes of algorithms, specifically designed for this emerging technology, as applied to signal processing, are derived.
Electronic neural networks made to learn faster by use of terminal teacher forcing. Method of supervised learning involves addition of teacher forcing functions to excitations fed as inputs to output neurons. Initially, teacher forcing functions are strong enough to force outputs to desired values; subsequently, these functions decay with time. When learning successfully completed, terminal teacher forcing vanishes, and dynamics or neural network become equivalent to those of conventional neural network. Simulated neural network with terminal teacher forcing learned to produce close approximation of circular trajectory in 400 iterations.
A method and apparatus for supervised neural learning of time dependent trajectories exploits the concepts of adjoint operators to enable computation of the gradient of an objective functional with respect to the various parameters of the network architecture in a highly efficient manner. Specifically, it combines the advantage of dramatic reductions in computational complexity inherent in adjoint methods with the ability to solve two adjoint systems of equations together forward in time. Not only is a large amount of computation and storage saved, but the handling of real-time applications becomes also possible. The invention has been applied it to two examples of representative complexity which have recently been analyzed in the open literature and demonstrated that a circular trajectory can be learned in approximately 200 iterations compared to the 12000 reported in the literature. A figure eight trajectory was achieved in under 500 iterations compared to 20000 previously required. The trajectories computed using our new method are much closer to the target trajectories than was reported in previous studies.
A neural network is trained to output a time dependent target vector defined over a predetermined time interval in response to a time dependent input vector defined over the same time interval by applying corresponding elements of the error vector, or difference between the target vector and the actual neuron output vector, to the inputs of corresponding output neurons of the network as corrective feedback. This feedback decreases the error and quickens the learning process, so that a much smaller number of training cycles are required to complete the learning process. A conventional gradient descent algorithm is employed to update the neural network parameters at the end of the predetermined time interval. The foregoing process is repeated in repetitive cycles until the actual output vector corresponds to the target vector. In the preferred embodiment, as the overall error of the neural network output decreasing during successive training cycles, the portion of the error fed back to the output neurons is decreased accordingly, allowing the network to learn with greater freedom from teacher forcing as the network parameters converge to their optimum values. The invention may also be used to train a neural network with stationary training and target vectors.
A neural algorithm for rapidly simulating a certain class of nonlinear wave phenomena using analog VLSI neural hardware is presented and applied to the Korteweg-de Vries partial differential equation. The corresponding neural architecture is obtained from a pseudospectral representation of the spatial dependence, along with a leap-frog scheme for the temporal evolution. Numerical simulations demonstrated the robustness of the proposed approach.
The thrust of this paper is to present a new approach to multi-target tracking for the mid-course stage of the Strategic Defense Initiative (SDI). This approach is based upon a continuum representation of a cluster of flying objects. We assume that the velocities of the flying objects can be embedded into a smooth velocity field. This assumption is based upon the impossibility of encounters in a high density cluster between the flying objects. Therefore, the problem is reduced to an identification of a moving continuum based upon consecutive time frame observations. In contradistinction to the previous approaches, here each target is considered as a center of a small continuous neighborhood subjected to a local-affine transformation, and therefore, the target trajectories do not mix. Obviously, their mixture in plane of sensor view is apparent. The approach is illustrated by an example.
A novel approach is presented which is motivated by an attempt to remove one of the most fundamental limitations of artificial neural networks: their rigid behavior as compared with even the simplest biological systems. It is demonstrated that non-Lipschitzian dynamics, based on the faliure of the Lipschitz conditions at repellers, displays a new qualitative effect, i.e., a multichoice response to periodic external excitations. This makes it possible to construct unpredictable systems, represented in the form of coupled activation and learning dynamical equations. It is shown that unpredictable systems can be controlled by sign strings which uniquely define the system behavior by specifying the direction of the motions at the critical points. Unpredictable systems driven by sign strings are extremely flexible and can serve as a powerful tool for complex pattern recognition.