Contributed articleReal time distributed processing of multiple associated pulse pattern sequences
Introduction
The world is full of sensors! Some of them, such as frequency spectrum analysers or thermal/visual imaging devices, give a complex temporal output signal which normally requires considerable interpretation. This in turn requires a good working knowledge of the general characteristics of both the sensor calibrations and operating principles as well as the phenomena being monitored. The output from a number of these more sophisticated sensors can be in the form of a temporal series of patterns transmitted as pulses along a parallel highway similar in form to the parallel ports on many pieces of electronic equipment. An artificial neural network is described and simulated which can simultaneously accept this form of raw data from more than one sensor and associate it directly with a well defined and stable pattern sequence which is easy to interpret, display, or can be used to directly control actuators.
The network described is a Real Time Distributed Associative Memory Artificial Neural Network (RTANN). This network was developed from an aid to pattern recognition described earlier (Travis, 1982, Travis, 1984). The present RTANN now has the ability to simultaneously handle more than three input pattern sequences and also has a greatly enhanced storage capacity. A version of the network has been used to analyse data from a video camera in order to actively control the transverse intensity distribution of laser light within an aperture (Travis, 1995).
Some artificial neural networks (Hopfield, 1982, Sobajic et al., 1988, Sandler et al., 1991) process a single input pattern represented by action potentials applied to its numerous parallel inputs and give an associated output pattern represented by fixed action potentials at its parallel outputs. Networks processing a temporal sequence of patterns (Kohonen et al., 1981, Wolf, 1988, Hertz et al., 1991) receive and produce time varying action potentials at its input and output connections. Within the present primitive RTANN action potentials are restricted to two values (0 and 1). Time varying action potentials to and from the RTANN thus take on the form of pulse trains.
The basic building block of the RTANN is the section pair. This will associate a single pair of pattern sequences. The symbol used to represent a section pair is shown in Fig. 1. The section pair is trained by presenting associated pattern sequences in real time, as indicated in Fig. 1. A pair of pattern sequences has only to be presented a single time to complete the training. Many different pairs of pattern sequences can be retained simultaneously in the section pair's memory. One of these sequences can then be regenerated simply by feeding its associated sequence into the other section (Fig. 2). If more storage space is required a number of section pairs may be stacked together to form a multi-layer section pair. To associate more than two pattern sequences section pairs are linked together.
The architectural flexibility allowed by the modular design of the RTANN allows networks to be configured for specific applications. An example is shown in Fig. 3. Here the network has been specifically designed to allow the output from two notional sensors monitoring the identity and colour of an object to be associated directly with a descriptive sentence having the form ‘These objects look colour.’ The simulation of this network is presented in Section 5.
Section snippets
General description of the RTANN
The RTANN is an assembly of many section pairs. A simple three bit section pair is shown in Fig. 4. The full compliment of connections to each of its units is shown in Fig. 5.
With reference to Fig. 4, the number of clusters of units within a section (Nc) is equal to the number of bits forming the input pattern. As indicated in Fig. 4, each unit within a cluster is connected exclusively to a unit in every cluster of the section by the local connections. With reference to Fig. 5, it will be
Functional elements
With reference to Fig. 6, Fig. 7, the characteristics of the functional elements (FEs) are as follows:
The training mode
With reference to Fig. 1, Fig. 4, Fig. 6, Fig. 7, the section pair is in the training mode whenever there are input pattern sequences to both sections. Even during the training mode there will always be some random pulse coincidences on Type 1/2 FE's which will cause an underlying component T1Lran of chemical messenger L within a cluster. At the same time Type 2 FEs are continually releasing sufficient chemical messenger K to maintain a background component T2Kbkd, whilst the crossover pulses
Simulation of an RTANN
The architecture of the simulated RTANN is shown in Fig. 3. It is an eight bit RTANN which will process sets of six associated input pattern sequences and is fabricated from eight six layer section pairs. The six pattern sequences represent the outputs from two notional sensors monitoring the shape and colour of an object as represented by a sentence having the form ‘These Objects look Colour’.
The 20 input sentences and sensor stimuli used during the network training phase are listed in
Conclusions
An artificial neural network has been described and its ability to simultaneously process a number of associated pattern sequences demonstrated. The Real Time Distributed Associative Memory Artificial Neural Network is the direct descendant of an earlier pattern recognition network (Travis, 1982, Travis, 1984) and incorporates a number of improvements.
Within the earlier network the control of unit activity and the triggering a cluster was governed by an external computation which took into
References (9)
- et al.
Statistical mechanics of temporal association in neural networks with transmission delays
Physical Review Letters
(1991) Neural networks and physical systems with emergent collective computational abilities
Proc. Natl. Acad. Sci. USA
(1982)- et al.
Parallel models of associative memory
- et al.
Use of a neural network to control an adaptive optics system for an astronomical telescope
Nature
(1991)