Fault Tolerance and Fault-tolerant computing is the ability to produce correct results even in the presence of faults, errors, or unexpected conditions. One of the advantages of neuralnetworks is that the processing ...
详细信息
We designed and trained a modified time-delay neuralnetwork (TDNN) to perform both automatic lipreading ("speech reading") in conjunction with acoustic speech recognition in order to improve recognition bot...
详细信息
In considering distributed adaptive routing schemes for large networks with dynamic topology, the need for an unconventional shortest path algorithm arises from the excessive computation overhead associated with repea...
详细信息
In considering distributed adaptive routing schemes for large networks with dynamic topology, the need for an unconventional shortest path algorithm arises from the excessive computation overhead associated with repeated path/distance calculations. The authors provide the design specifics of such an algorithm, and establish its performance characteristics through rigorous analysis and simulation. The new algorithm exploits the intrinsic parallelism of neuralnetwork architectures and solves the single-pair shortest path problem in such a way that: (i) the computation time is independent of the number of network nodes; and (ii) the frequent shortest distance/path re-calculations inherently associated with topology changes are performed much faster than conventional algorithms. Simple exploitation of the inherent parallelism allows extension of the algorithm to solving single-source and all-pair shortest path problems without compromising the trait of constant convergent complexity.< >
The effects of the adjustment of the threshold of the hidden cells during learning in a one-hidden-layer backpropagation network with half-distributed coding of inputs are analyzed. The fundamentals of this coding met...
详细信息
The effects of the adjustment of the threshold of the hidden cells during learning in a one-hidden-layer backpropagation network with half-distributed coding of inputs are analyzed. The fundamentals of this coding method are reviewed. Although it can be applied to both inputs and outputs of the network, only the case of the inputs is considered. The effects of the modification of the thresholds during learning are analyzed. It is shown that these effects become more favorable as the task to be achieved becomes less complex. The correctness of the theoretical model was tested with a real-world application. The network has to approximate a function to realize a numerical model of a physical phenomenon.< >
This study demonstrates a paradigm for modeling speech production based on neuralnetworks. Using physiological data from speech utterances, a neuralnetwork learns the forward dynamics relating motor commands to musc...
ISBN:
(纸本)9781558602748
This study demonstrates a paradigm for modeling speech production based on neuralnetworks. Using physiological data from speech utterances, a neuralnetwork learns the forward dynamics relating motor commands to muscles and the ensuing articulator behavior that allows articulator trajectories to be generated from motor commands constrained by phoneme input strings and global performance parameters. From these movement trajectories, a second neuralnetwork generates PARCOR parameters that are then used to synthesize the speech acoustics.
This paper uses the Backpropagation algorithm as an example to discusses different strategies of implementing neuralnetwork,algorithms on the massively parallel computer of the Connection Machine *** most of these al...
详细信息
This paper uses the Backpropagation algorithm as an example to discusses different strategies of implementing neuralnetwork,algorithms on the massively parallel computer of the Connection Machine *** most of these algorithms involve only local computations at each processing unit that can be carried out in parallel,the neuralnetworks are usually highly *** careful design,an implementation of such algorithms on a massively paraEel,distributed memory computer can easily spend the majority of its running time in communications rather than in actual *** major motivation for our work was exactly to avoid this *** a 32K processor CM-2,the best performance of 150 million Weight Update Per Second(WUPS) was obtained through a data parallel implementation *** techniques described in this paper could serve as examples for implementing other neuralnetwork algorithms on any massively parallel,distributed memory computers that have general communication architectures.
With the development of anditory models,many methods of explaining a lot of perceptual pitch phenomens such as "pitch of the missing fundamental" and simulating many aspects of human pitch perception by a pe...
详细信息
With the development of anditory models,many methods of explaining a lot of perceptual pitch phenomens such as "pitch of the missing fundamental" and simulating many aspects of human pitch perception by a peripheral auditory model were proposed. These methods were fundamentally based oa the summary autocorrelation processing,which simulated the spatial integration function of the human auditory system,and the central neural system detected pitch information included in the integration *** these methods could explain lots of perceptual pitch phenomena and extract successfully the pitch information from the pare speech signal or the signal deteriorated by noise,it was not able to detect effectively two individual pitch information included in the overlapping speech signal because of the summary masking influence of the stronger signal.
＜正＞This paper introduced the attributed relation graph to represent 3-D object and applying Hopfield network to recognize object based on three *** result turns out that this approach needs less preprocessing and g...
＜正＞This paper introduced the attributed relation graph to represent 3-D object and applying Hopfield network to recognize object based on three *** result turns out that this approach needs less preprocessing and gives out more convenient and effective method to construct the attributem relation *** parallel distributedprocessing mechanism of Hopfield metwork makes it possible to speed up the process of *** convenient sake,In the study of nibostness of the method,we choose more complicated "face image" in addition to machine part image as *** result is perfect.
The following topics are dealt with: the European Synchrotron Radiation Facility (ESRF) in Grenoble, France;neuralnetworks in front-end processing and control;a neuralnetwork in expert systems;FASTBUS network initia...
详细信息
The following topics are dealt with: the European Synchrotron Radiation Facility (ESRF) in Grenoble, France;neuralnetworks in front-end processing and control;a neuralnetwork in expert systems;FASTBUS network initialization;hypercard graphic interface toward CAMAC and VME;the VME intercrate bus;data acquisition architectures;dynamic debugger for multiprocessor real-time applications;data acquisition for the Superconducting Super Collider;an application of the PS1 TANDEM system;distributed control system for NAC cyclotrons;MEMPHIS 2000, a modular experiment multiparameter pulse height instrumentation system;a VMEbus clock system for accelerator control;automatic alignment of a Michelson interferometer;and GOOSY-VME, hardware components and trigger system. Abstracts of individual papers can be found under the relevant classification codes in this or other issues.
A novel keyword-spotting system that combines both neuralnetwork and dynamic programming techniques is presented. This system makes use of the strengths of time delay neuralnetworks (TDNNs), which include strong gen...
详细信息
A novel keyword-spotting system that combines both neuralnetwork and dynamic programming techniques is presented. This system makes use of the strengths of time delay neuralnetworks (TDNNs), which include strong generalization ability, potential for parallel implementations, robustness to noise, and time shift invariant learning. Dynamic programming models are used by this system because they have the useful capability of time warping input speech patterns. This system was trained and tested on the Stonehenge Road Rally database, which is a 20-keyword-vocabulary, speaker-independent, continuous-speech corpus. Currently, this system performs at a figure of merit (FOM) rate of 82.5%. FOM is the detection rate averaged from 0 to 10 false alarms per keyword hour. This measure is explained in detail.< >
暂无评论