Clark S. Lindsey

Survey of neural network hardware

We survey the currently available neural network hardware, including VLSI chips (digital, analog, and hybrid), PC accelerator cards, and multi-board neurocomputers. We concentrate on commercial hardware, but also include a few prototypes of special interest. As examples of applications, some systems developed for high energy physics experiments that use this hardware are presented.

Intelligent detectors modelled from the cat's eye

Abstract Biologically inspired image/signal processing, in particular neural networks like the Pulse-Coupled Neural Network (PCNN), are revisited. Their use with high granularity high-energy physics detectors, as well as optical sensing devices, for filtering, de-noising, segmentation, object isolation and edge detection is discussed.

Method for star identification using neural networks

Identification of star constellations with an onboard star tracker provides the highest precision of all attitude determination techniques for spacecraft. A method for identification of star constellations inspired by neural network (NNW) techniques is presented. It compares feature vectors derived from histograms of distances to multiple stars around the unknown star. The NNW method appears most robust with respect to position noise and would require a smaller database than conventional methods, especially for small fields of view. The neural network method is quite slow when performed on a sequential (serial) processor, but would provide very high speed if implemented in special hardware. Such hardware solutions could also yield lower low weight and low power consumption, both important features for small satellites.

Filtering data from a drift-chamber detector using neural networks

Abstract This paper describes the implementation of an analog neural network chip (Intel 80170NX ETANN) to filter a multi-signal read-out from flash-ADCs used in a liquid argon time projection chamber (ICARUS). The training and subsequent testing of the network is discussed in some detail. A network consisting of multiple sub-networks is proposed to perform signal filtering, determination of peak position and track finding.

Hybrid neural networks for automatic target recognition

The paper presents a hybrid neural network system for automatic target recognition, or ATR. The ATR system uses a hybrid of a biological inspired neural net called the Pulse Coupled Neural Net, PCNN, and traditional feedforward neural nets. The PCNN is an iterative neural network in which, for example, a grey scale input image results in a 1D time signal invariant to rotation, scale and translation alternations. The PCNN can also extract edges, perform object segmentation and extract texture information. The PCNN pre-processor generates a 1D time signal that is input to a feedforward pattern recognition net.

Implementing the dynamic decay adjustment algorithm in a CNAPS parallel computer system

Abstract This paper presents the implementation of the Dynamic Decay Adjustment (DDA) algorithm in a CNAPS parallel computer system having 128 processing nodes. The DDA algorithm has several inherent advantages, and the implementation of it in the CNAPS system is shown to perform very well. The DDA implementation is first tested with noisy character patterns to demonstrate its general inherent noise resistance. A more realistic application test involving the identification of Higgs events is then presented. Using the momenta and transverse momenta of the four leading particles from the H 0 → Z 0 Z 0 → μ + μ − μ + μ − decay following gg and W + W − fusion, it is possible to obtain a good indentification of these events as well as good rejection of the background.

Wavelets and signal processing

We briefly review the use of the wavelet and wavelet packet transforms. We describe their application to signal processing as devices for feature extraction and reduction of data for neural networks, in particular their implementation in hardware for signal identification.

Implementing the new Zero Instruction Set Computer (ZISC036) from IBM for a Higgs search

Abstract Implementation of the new IBM Zero Instruction Set Computer (ZISC036) on a PC/ISA-bus card as well as on a VME-card is reported. The ZISC circuit has 36 processing elements of a type similar to that of Radial Basis Function (RBF) neurons. It is a highly parallel and cascadable building block with on-chip learning capability, and is well suited for pattern recognition, signal processing, etc. Results of a test on identification of simulated Higgs events are given.

Investigation of a VLSI neural network chip as part of a secondary vertex trigger

Abstract An analog VLSI neural network chip (ETANN) has been trained to detect secondary vertices in simulated data for a fixed target heavy flavour production experiment. The detector response and associative memory track finding were modelled by a simulation, but the vertex detection was performed in hardware by the neural network chip and requires only a few microseconds per event. The chip correctly tags 30% of the heavy flavour events while rejecting 99% of the background, and is thus well adapted for secondary vertex triggering applications. A general purpose VME module for interfacing the ETANN to experiments, equipped with ADC/DAC circuits and a 68070 CPU, is also presented.

Unsupervised learning with ART on the CNAPS

Abstract The most popular types of neural network training, such as back-propagation, involve a teacher, or supervisor, that provides a desired output for each input pattern. Supervised algorithms then attempt to minimize differences between the network outputs and the desired outputs. Unsupervised training algorithms, on the other hand, attempt to categorize the data without any external guide to the class of a given training pattern. Such networks can be valuable for analyzing data to search without bias for unknown classes, trends and relationships. Adaptive Resonance Theory (ART) algorithms, in particular, are a very popular form of unsupervised training. ART networks can not only learn to differentiate data into categories, but they easily learn new classes without destroying prior learning (solving the so called “stability-plasticity” dilemma). Here we report on the implementation of ART networks on the Adaptive Solutions CNAPS parallel processor system to obtain very fast unsupervised learning. We aim to use these capabilities for applications in “data mining” of large data sets such as those from high energy physics, remote sensing, etc.