The Photoneuron - A Dynamically Reconfigurable Information Processing Control Element utilizing Embedded Fiber Waveguide Interconnects Andrew S. Glista Jr. Naval Air Systems Command Code AIR 546-TE Washington, DC 20361 ABSTRACT The term "photoneuron" describes an electro-optic hardware element which permits an optical implementation of the postulated information transfer processes of the neurons in the human brain. The photoneuron provides a dynamic activation and control mechanism for highly parallel computers and permits immediate implementation of reconfigurable high speed optical interconnects. The suggested method for interconnecting processors in a photoneuronic network consists of embedded optical fibers in composite materials to form optical backplanes utilizing "smart skin" technology. This method eliminates the environmental concerns and technological barriers posed by free space optics and integrated optics, while providing a sound engineering approach leading to the all optical computer. This paper briefly reviews the physiological activity of neurons in the human brain. Optical analogies for processor activation in neural networks corresponding to the nerve impulse activation in the brain are then described. The paper then suggests the utilization of optical signal parameters and encoding to emulate the information exchange of neurotransmitters provided by first and second messenger molecular activity across the synaptic "connections" of neurons in the brain. This represents a departure from most neural networks which dwell on threshold processor activation and ignore the exceedingly complex molecular information exchange mechanisms of the brain. Digital, analog, and combinatorial a1ternatives are described. 2. NEURAL PROCESSING 2.1 Physiological models The human neuron, a cell a few microns in diameter, is the basic building block of the human brain and nervous system. The human brain is estimated to contain approximately 1011 neurons. Each neuron possesses its own energy source in the form of mitochondria, cells which metabolize glucose and other food substances to manufacture biological energy. A single fiber connection called an axon extends from each neuron. These axons can be as long as three feet in length and, when activated, send an electrical pulse down the length of the fiber. These pulses can travel at rates of from one to one hundred and fifty miles per hour down the axon as the result of ionic exchange between sodium and chloride ions across the axon fiber membrane. The axon separates into small fibers, each of which serve as a connection to other neurons. This "connection" is called a synapse and is not really a connection in the physical sense since no physical contact is present. Rather, a gap is formed at each synapse between the axon fiber end and a receiving fiber on another neuron. These receiving fibers called dendrites are quite numerous as contrasted with the single axon sending information and serve as receptors of information from other neurons via the synaptic connections. The possibilities for interaction considering all combinations and permutations within a system with such a staggering number of elements approach infinity. The transfer of information across the synapse from the axon to the dendrite is not made via the electrical pulse or sequence of pulses which are sent down the axon. The pulse rather serves to release a chemical shower across the synaptic "connection" (actually a space). It is in this chemical release that information is transferred from the axon to the dendritic receivers of other neurons. These chemical neurotransmitters can be either excitory or inhibitory. An excitory neurotransmitter will contribute to the charge buildup in the receptor neuron while the inhibitory neurotransmitter will negate the charge buildup in the receptor neuron. The total information entering a neuron will either cause the neuron to "fire" (i.e. send out an electrical pulse causing the transmission of information to dendritic receivers of other neurons across the synaptic gap) or to remain idle. The information transfer mechanism between neurons in the brain is theorized to be contained in the chemical release across the synaptic connection. While the information transfer molecule serves as the messenger, it does not necessarily contain the entire message (i.e. the information). When messenger molecules are transmitted across the synaptic gap and attach to their receptor molecules on the receiving neuron, they may trigger the neuron into action by changing the electrolytic potential causing the above described electrical pulse to be generated and transmitted down the axon. other receptor molecules, part of a second-messenger system, act indirectly after being triggered by the first messenger molecule at the synapse. The electrical pulse can cause the body organs to produce hormones or enzymes which in turn can act on other neurons via the blood supply. This second messenger system can actually act on DNA to cause long-lasting or perhaps permanent changes in the cells. Some psychobiologists speculate that this might be the physical phenomena associated with long term memory. The information transfer, including memory, may be a combination of a complex chemical information transfer and storage process and the connectionist activity between the neurons. The shape of the molecules of the chemical "shower" has been shown to be a factor in the information transfer across a synapse in experiments involving morphine related drugs where different substances with similar molecular shapes produced similar effects. A "lock and key" mechanism based on the molecular structure of the synaptic connection is theorized to provide the activation or inhibitory response in the individual neuron since many unrelated molecules of similar molecular structure result in similar effects. Theoretical understanding of the exact information transfer process and memory retention means of the human brain are highly speculative at this time, but highly complex electro-chemical and bio-chemical relationships based upon molecular structure are involved. The enormous number of elements (both neurons and interconnects) in the brain provides for fault tolerance in the event that individual elements are damaged. However, understanding of the exact operation of this fault tolerance mechanism currently remains hypothetical at best.1,2 2.2 Electronic Artificial Neural Networks The development of artificial neural networks is an attempt to emulate the architecture of the human brain through electronic means. The current neural network concepts being explored are based on concepts and knowledge (both electronic and physiological) developed in the pre-digital computer era. Current electronic neural networks are comprised of the following elements: • a multiplicity of processing elements (neurons) producing a binary quantized output. • a fanout mechanism and interconnect means from the output of the processing elements (neurons) to a multiplicity of other neurons. • a means for implementing a multiplicity of data inputs to each neuron. • a means of applying adaptive coefficients or weights to each input • a means of summing the weighted inputs to each neuron. • a means for activating the processing elements by comparing the summation of the weighted inputs to a given threshold level according to a prescribed linear or non-linear activation function (usually a sigmoid function). • a feedback or feedforward mechanism to change coefficients or weights in the inputs or to serve as an input to other neurons. • an optional limited memory retention means. Figure 1 is a schematic diagram of such a neural network. Many of these "networks" utilize relatively simple analog processing elements (i.e.operational amplifiers). A large number of these elements are interconnected in a predetermined manner. The activation is usually accomplished by summing the analog inputs. If the summation of the input voltages exceeds the activation threshold (determined by a sigmoid function), the element begins processing. The current neural network "processor" usually consists of a one bit analog to digital converter. The inputs to a neuron can be "weighted" in a positive or negative sense which is analogous to the excitory or inhibitory action of charge buildup in biological neurotransmitters. Feedback can be utilized to adjust these weights and thus effect "learning". Once a neuron is activated and processes the incoming data signals, theoutput signal is sent to a predetermined number of related processing elements through a fanout mechanism and in turn these signals can be weighted and/or summed providing the stimulus for initiating the processing at other nodes. Weights can be adjusted dynamically through feedback or feedforward loops. The entire processing network is driven as a chain reaction by the dynamic input signals. The output is determined by the "hardwired" interconnect system and the processing power of each element. This type of processing can require a massive number of interconnects and processing elements and "software" is not extensive as in von Neumann machines. This arises from the fact that the "built-in" architecture and dynamically alterable input signal activation serve as the programming mechanism. It should be noted that most activity to date in the implementation of artificial neural networks has been centered around the "activation process". This summation and threshold activation process described above corresponds to the process in the human brain which is responsible for sending the nerve impulse or activation potential down the axon and eventually causes the information flow by neurotransmitters across the synaptic gap. The activation process typically generates a binary decision process. This simplistic "processing" is not at all representative of the complex electro-chemical and biochemical activity occurring within the neuron which results in the transfer of information at the human synaptic interface. Neural Networks, as presently implemented, ignore the complex first and second messenger systems associated with complex molecular coding and activation and present a rather weak implementation of fault tolerance based on merely ignoring inputs which cause a change in the threshold activation energy. The ability to implement neural networks and analog and digital computation via electronic signals is severely limited by several factors: • the capability for fanout is limited to a small number of nodes • data transmission rates are can be limited by electrical impedance because of resistance, capacitance, and power limitations. • no adequate electrical fault tolerant electrical interconnect means exists. • electrical interconnect density is limited by noise or crosstalk between adjacent input/output lines which can degrade signal to noise ratio. • low level electrical signals are susceptible to noise and electromagnetic interference (EMI). • achievement of uniform weighting resistances in neural network circuits with large physical dimensions is difficult. • most neural networks utilize each neuron to make a binary decision rather than to activate a more complex transform process. When done on a pixel by pixel basis as is the case for image recognition, an extremely high number of neurons is required since each neuron represents a one bit analog to digital convertor. Most existing electron neural networks are actually digital emulation of the activation process of neurons in the human brain on a conventional Von Neumann digital computer. No attempt has been made to date to duplicate the complex information processing and transfer functions associated with the first and second messenger systems of the brain achieved through molecular coding. Fault tolerance has been achieved by "ignoring" a failed input in the thresholding process rather than by pursuing a dynamically reconfigurable or "self healing" interconnect system which might more accurately reflect the physiological model. 2.3 Current Optical Interconnect Limitations Optical interconnection of the processors in neural networks have been widely investigated to overcome the above noted limits of electrical interconnects. The use of free space optical communication with holographic elements or spatial light modulators serving as distribution or fanout elements have been proposed. Spatial light modulator technology and alternative light deflection technology are currently limiting free space optical interconnects but these are not the only impediments associated with free space optical approaches. operation in harsh environments (especially vibration, shock, and temperature extremes) is another concern with this approach Geometric considerations limit the fanout potential of this approach. Even though free space optical beams can pass through each other without interference, these beams cannot pass through solid elements (i.e.photodetector arrays). Vibration or imperfections in optical deflection hardware can cause crosstalk between adjacent high density photodetectors. It becomes very difficult to implement multilayer neural networks or analog control circuits with free space optics where feedforward and feedback optical signals are required to impinge on a large number of physically separated receptors. Integrated optics and planar waveguide technology has received much attention for use in optical interconnect and computation. These waveguides exhibit high attenuation (typically on the order of .1-1 dB/cm versus .1-1 dB/km for commercial optical fibers) and are temperature sensitive. Interconnection to other substrates using free space optics or optical fibers is extremely difficult. Existing couplers and switches require large areas of real estate to obtain high fanout. 2.4 Fault Tolerant Optical Fiber Networks Reference 3 described a technique for achieving an extremely high degree of fault tolerance in fiber optic networks. An example of such a network is shown in Figure 2 where multiple nodes can be bypassed with uninterrupted network operation. Also described in reference 3 is a multidimensional high density programmable optical interconnect methodology for massively parallel computers such as hypercubes. Figure 3 shows the electro-optic elements of such a computer node consisting of multiple photodetectors, a means of comparing the input signals to a given threshold via analog comparators, a switching means to select a given input for Reference 3 described a technique for achieving an extremely high degree of fault tolerance in fiber optic networks. An example of such a network. is shown in Figure 2 where multiple nodes can be bypassed with uninterrupted network operation. Also described in reference 3 is a multidimensional high density programmable optical interconnect methodology for massively parallel computers such as hypercubes. Figure 3 shows the electro-optic elements of such a computer node consisting of multiple photodetectors, a means of comparing the input signals to a given threshold via analog comparators, a switching means to select a given input for retransmission, a laser or LED to provide signal amplification and a lXN optical coupler which splits the incoming primary input signal into feedforward or feedback bypass fibers which provides any desired degree of fault tolerance or connectivity in a network. 3.0 OPTICAL WAVEGUIDE NEURAL NETWQRKS While electronic artificial neural networks provide an analogy to the nerve impulse activation and transmission of the human brain, electro-optic technology can be utilized to more closely match the other postulated functions occurring in the brain including synaptic connections and first and second order messenger interchanges or self healing. These alternatives which will be described below can be utilized to dynamically activate or control analog and digital computer processing and memory functions, to implement neural networks, and to perform various processing functions including analog optical computation. 3.1 Fiber Optic Neural Networks The following section describes the implementation of neural networks with fault tolerant optical waveguide interconnects and related electro-optic elements. The combination of these electro-optic (or totally optical) elements in a terminal performs the activation function corresponding to the nerve impulse generation of the brain. Data signals output from a given processing element are converted to an optical signal through the utilization of an optical source such as a laser or LED or laser or LED arrays (See figure 4). The optical signal is coupled to a single optical waveguide from the modulated laser or LED or to multiple waveguides from the modulated laser diode arrays or LED arrays. These electro-optic sources may produce output with single or multiple wavelengths. If a laser or LED array is utilized, the array, whose elements are all modulated with the same signal, also serves as a fanout mechanism since each element of the array sends the same optical output signal to other processing elements. As is the case in an electrical neural network, a fanout mechanism is required to transmit the output to succeeding layers in the optical waveguide network. As mentioned above, the independently modulated elements of a LED or laser array can serve this purpose. In addition, a 1XN optical waveguide coupler can act as a means of providing even greater fanout from the output LEDs or lasers with N waveguides proceeding to higher or lower layers in the network from each element in the array. A biconical twisted fiber fusion coupler, a graded index glass coupler or a large core fiber to multiple fiber bundle can be utilized for this purpose and many concepts are well known in the art. 1XN couplers emanating from each element of the array can provide even greater fanout capability if desired (See figure 4). The fanout waveguides serve as input to the next layer in the neural network or can serve as a bypass or feedforward mechanism to a higher layer in a multilayer neural network hierarchy (see figure 5). The use of telecommunications fibers having a 125 micrometer outer diameter can produce an interconnect density of 40,000 interconnects per square inch when stacked in a two dimensional array. Even higher interconnect density can be achieved with small diameter waveguides such as those utilized in imaging fiber optic bundles where fibers with less than 20 micrometer diameters are used. This is mentioned to note that the density of interconnects that can be achieved with existing technology in implementing extremely high density neural network interconnects is extremely high. Combined with the extremely high bandwidth of each fiber, the figure of merit for a fiber waveguide neural network (in terms of interconnect speed-density) can approach that of the physiological model. Any planar optical waveguide technology can be utilized as the interconnect means. optical waveguides can be placed in extremely close proximity due to their EMI immunity but a real "rats nest" of optical wiring is apparent from the above diagrams. 3.2 High Density Optical Backplanes Utilizing Smart Structures Technology Smart structures" is the branch of materials science which deals with the use of embedded sensors within a material to monitor and control its mechanical properties. A great deal of attention has been given to embedding optical fibers in composite materials for both sensing and data transmission. This technology can be exploited to provide a monolithic high density two or three dimensional optical waveguide backplane to provide a structurally rigid packaging scheme for the massive number of interconnects which may be required in a neural network system. The combination of very low cost fiberglass structural fibers, optical fibers, embedded couplers and Silicon V-Groove connectors would provide an ideal combination of materials for this application. This monolithic structure containing the embedded fibers an/or couplers can serve to form an optical backplane when connectorized. (See Figure 6) In such a configuration, the processing elements and related summing and activation means can be packaged in a modular format. Glass/epoxy composite or any other composite containing the optical waveguides and any electrical or ground wires can also be utilized. High density connectors such as silicon V-groove which is currently well known in the art, coherent bundles of optical waveguides, or fused optic faceplates waveguides and connectors (See figure 7) can be utilized to couple to the processing elements input photodetectors and output laser/LED arrays in a modular fashion. The fibers, couplers, and connectors can be suitably embedded in the optical backplane substrate material to form a rugged monolithic packaging scheme. Couplers with hundreds of outputs from each laser array element are possible. The multiple waveguides from each node with the capability to bypass any number of failed processor nodes provides a level of fault tolerance which is difficult to achieve with electrical interconnects or free space optics. These high density fanout waveguides can serve as a feedback mechanism to lower layers in the network to adjust input weights such as an X-Y addressable controlled transparency. (See Figure 8) Frequency selective couplers can also be utilized to determine fanout paths if desired. In an optical waveguide implementation of a neural network, the output of a processor (neuron) is converted to optical format through a LED or laser diode array. The optical fanout waveguides are incident on a large area photodetector, multiple photodetectors or photodetector arrays which serve as receptors of the incident optical signal in the next layer of the network. The individual photodetectors may be PIN photodiodes, avalanche photodiodes, phototransistors, or charge coupled devices (CCDs) or arrays of such devices. These photodetectors can also be wavelength selective if desired. As in the case of electrical networks, a means of providing weighting to the individual inputs is required. The information transferred over the fibers can be analog or digital and the weighting can be implemented in the electrical or optical domain through optical signal modulation. In an analog implementation of a fiber optic neural network, the weights to the individual optical waveguide inputs may be provided in several methods. One such implementation of weighting in the electrical domain utilizes Metal Oxide Semiconductor (MOS) transistors as variable resistors in series with each photodetector. The current output from the photodetectors flowing through the channel of the transistor can be adjusted in amplitude by varying the gate voltage. The values of the gate voltage to be applied to the weighting transistors can be sent down the optical waveguide from other nodes as a signal superimposed on the data signal. The weighting information can also be sent as another frequency signal down any of the input waveguides as well. A method of weighting the signals of the optical waveguide neural network element in the optical domain is by individually controlling the amplitude of the optical output of each element by regulating the amplitude of the drive current through the individual light emitting diodes or semiconductor lasers in an array to provide inhibitory weighting. Transistors or variable resistors can be utilized to control the amplitude and can be fabricated as part of the array or as part of the control circuitry driving the elements of the array. (See Figure 9) Low threshold, individually addressable quantum well surface emitting laser diode arrays are currently under development. This method of weighting would not require couplers for the fanout mechanism since the LED or laser diode array would serve this purpose. In the event that 1XN couplers are utilized to obtain additional fanout, (See Figure 10) an alternative method of weighting on individual output fibers of the coupler would be required to assure independent weighting. Many alternative methods exist for the weighting the signals of an analog neural network interconnected with optical waveguides in the optical domain. They include the use of a suitable material which can vary the optical transmission from the waveguide to the photodetector in response to an electrical or optical control signal. This would mimic the inhibitory response of the human neuron. These modulator materials such as a photochromic, electrochromic, or liquid crystal can be in the form of an x-y addressable window placed in front of the photodetectors or laser diodes or can even be a layer of suitable material deposited directly on the devices. controlling the electrical or optical bias on the variable transparency material will serve as the means for controlling the weights. Another method of weighting the signals input to the processing element of an optical waveguide neural network in the optical domain is through the use of optical amplifiers to boost the signal in a given optical waveguide proportional to the desired weight. This additive (excitory) weighting can also be used to overcome losses in the waveguide transmission media where alternative weighting concepts are implemented. Just as the input signal weighting can be accomplished by various electro-optic techniques, the means for summing the weighted signals in an optical waveguide neural network or processor activation may be accomplished with several techniques in the electrical or optical domain: A method for summing the weighted outputs is through the use of a large area photodetector. (See Figures 9 and 10) This method is possible if the weighting is accomplished by amplitude control of the individual elements of the optical source LED or laser diode array, or by the use of controlled transmission or absorption materials. Another summing method uses a graded index lens or solid glass mixing rod wherein the optical modes are mixed and then fed to the photodetector. This technique utilizes summation in the optical domain and the output of the photodetector converts the summed signal to the electrical domain. The threshold activation function of the optical waveguide neural network can be accomplished in the electrical domain after the optical signal has been converted to electrical form by a photodetector. This activation is accomplished with analog electronic comparators wherein the summation of the weighted electrical signals is compared to a fixed or programmable reference voltage. The function that determines the activation threshold may be linear or non-linear and correlators may also be used for this purpose. In the optical domain, optical bistable elements including symmetric self actuated electrooptic effect devices (S-SEEDs) can be used for threshold activation wherein the optical element changes state based upon the amplitude of the optical power in the input waveguides. The parameters of the bistable element or combinations of elements can be varied to set the threshold level. 4.0 BEYOND NEURAL NETWORKS 4.1 Dynamically Reconfiqured Interconnects The previously described optical waveguide neural networks are similar to electronic neural networks which mimic the ionic activation process of the brain and control a binary decision at each neuron. A notable difference between the two is that the use of waveguide and electrooptic technology permits a direct hardware implementation of neural networks rather than a simulation of the network on a von Neumann computer. The same dynamic activation methods utilizing optical waveguides can also be extended to reconfigure interconnects between computer processing elements. A fanout consisting of a lXN coupler emanating from a laser or LED or multiple couplers emanating from laser or LED arrays serves to send information to a large number of processing elements. The optical properties of the information on these fibers can control switching devices which can vary the receptors of the information on the lines. An example of this is the use of wavelength sensitive couplers in the networks so that information of different wavelengths is routed to different processing nodes. Active devices can also be used to implement real time circuit switching of information. An optical crossbar switch for dynamic signal routing can be constructed with electrooptic elements and can be changed or set by an optical input signal. The amplitude, frequency, modulation format, or polarization of an incoming signal can be utilized to activate the switching of the elements in an array of receiving photodiodes in a processing element from an off to on state. Since the output of any processing element is connected to the input of all desired processing elements via the fanout mechanism, the only processors which will be connected are those whose receiving device is switched into the circuit via the logic activated by signal amplitude, frequency, modulation format, phase, or polarization. Frequency selective couplers and optical bistable elements could be used for the optical implementation.(See Figure 11) 4.2 Optical Activation Of Information Transform Processors Neural networks use the thresholding of summed weighted inputs to achieve activation in a neural network. Dynamic adaptive activation of complex computer elements and signal processors can also be achieved by altering the parameters of light (intensity, frequency, modulation format, phase or polarization) passing through high density optical fiber waveguide interconnects. Controlled variation of these optical parameters can be used to activate and de-activate processing elements, alter and control transfer functions or processing sequences, control and change input/output functions, and to control and alter the storage and retrieval of information from memory. Such a collection of electro-optic elements can an be characterized as an electro-optic "photoneuron" and is shown in Figure 12. 'The processing elements of a photoneuronic network, which can perform complex transforms as contrasted with the binary decision process of a typical neural network "processor", more accurately reflects the complex information transformation processes occurring in the human neuron vice the binary activity occurring in the ionic activation process. As such, these processing elements can perform complex information transformations which can be implemented in either analog or electronic digital integrated circuit format, totally optical format or a combination of the above. This processing method can be technologically transparent to the optical transmission format, the activation process or the weighting process. Activation of these processors could be accomplished in a fashion similar to the molecular lock and key activation associated with the complex first and second order messenger signal transfer process occurring at the synaptic junction. A simple amplitude controlled activation technique could be implemented by varying the amplitude of digitally encoded data being transferred through optical waveguide computer interconnects to activate selected processors. A fanout consisting of a lXN coupler emanating from a laser or LED or multiple couplers emanating from laser or LED arrays would serve to send information to a large number of processing elements. The relative amplitude of the digital pulse train would control which processors would be activated. The level could be determined by adjusting the laser diode array elements intensity or through the X-Y addressable modulators. Peak level detectors would determine the amplitude of the optically encoded digital pulses and thus activate the processing element. The pulse train itself would be the information to be acted upon by the processor. An alternative implementation would utilize threshold logic circuits to perform a different set of operations based on the peak level of the of the incoming signals. Alternatively, the wavelength or polarization of the incoming signal can be utilized to selectively activate a predetermined processing sequence. As is the case with conventional analog computers, operational amplifiers, and a combination of resistive, inductive and capacitive elements, can be hardwired to implement the desired transfer functions. These functions can include adders, multipliers, dividers, differentiators, integrators, etc. The electronic analog transfer functions can be altered in real time within a given network node by varying the resistance, capacitance or inductance of the circuit. The optical parameters of wavelength, amplitude, modulation format or polarization of the incoming signal would determine which transfer function would be selected by a switching means. Resistive ladders, variable resistors, switched capacitor banks, or tapped inductors can be utilized as variable circuit elements which could be reconfigured by switching.This method obviates many noise, impedance, and accuracy limitations of previous electronic analog computers by eliminating crosstalk, electronic patch cords, and electronic "tweaking" since the information is in the optical domain. Both electronic or optical switching devices could be utilized. Optical signals can also be utilized to directly vary the resistance, capacitance, or inductance of the control circuit elements with the proper selection of bandgap engineered devices utilized for this purpose. The result of the analog processor transform operation would determine the characteristics of the optical output signal parameters from the node. This output would in turn activate and control the processing transfer functions of successive (or prior) layers of the network through feedback or feedforward connections. Complex digital processing elements can also be activated in accordance with a fixed or dynamically programmed binary or N-ary code. The optical signals transmitted by a high density array of optical fibers incident on the X-Y matrix of photodetectors can serve to activate various stored programs within the processor in response to the incident X-Y light pattern. This would also be analogous to the "lock and key" activation arrangement postulated for the human neuron. The light intensity pattern incident on a photodetector array can be utilized to select which processor operations can be triggered, which stored program should be activated, or which memory should be accessed. B changing light intensity, comparator threshold settings, and X-Y address coordinate settings, and fantastically high combination of associations can be made in a small photodetector array physical The use location. of digital or analog correlators can also be used for this "lock and key" activation. This same combination of electro optic elements can be used as an optically activated threshold memory. It can be constructed with an input light source, an X-Y addressable matrix of variably transmissive optical materials, an array of photodetectors, and an array of comparators/correlators. (See figure 13) To store information in this memory, The X-Y lines of the variably transmissive material are activated corresponding to the desired stored binary pattern (i.e. a transmissive element represents a "1" and opaque element represents a "0"). The threshold of the comparators is set to a level which will produce an output only when the light transmitted through the transmissive element striking the photodetector is of sufficient intensity. These areas can be interpreted as binary logic level "ls". The transmission through the opaque elements will not possess sufficient intensity to produce photocurrent necessary to produce output from the threshold comparator since they are below threshold. These areas can be interpreted as a binary logic level "Os". To produce non-volatile storage of a complex array pattern, only coefficients of the X-Y transparency matrix, the light intensity level, and the comparator threshold levels need be preserved to restore the entire X-Y bit pattern. Since the comparator threshold levels and light intensity are variable, a given physical location can store multiple or layered memory patterns. If the comparator threshold level is now raised above the previous level required to produce photodetector output from the illuminated devices, the previously set light intensity will produce no output from any of the comparators. Thus a clean slate is provided for storage of a new memory pattern in the same physical location by resetting the light intensity and the comparator, and the XY addresses of the transparency matrix. Such a scheme can be utilized as an optical associative memory similar to the lock and key analogy of the human brain. The use of various image algebra processing and/or correlation schemes can then be used to access such an associative memory from fuzzy or incomplete optical input data. This technique can also be utilized to provide multistate (threshold or fuzzy) logic processing elements rather than binary logic elements if desired. The encoding can be "N-ary" as contrasted with the binary logic methods currently used. The low loss and low noise properties of optical transmission thus make possible alternative techniques for implementing optical threshold or fuzzy logic. Figure 14 is a corresponding "photoneuron " with all optical elements and utilizes optical summation, activation and processing elements so that conversion of the signal to electronic format is not required. S-seed type devices have been previously described in the literature for this type of digital activation. The means is currently available to produce all elements of a totally optical waveguide analog computer. Such a method utilizes optical sources, amplifiers, attenuators, modulators and storage/branching elements to alter the parameters of light traveling through optical waveguides directly analogous to the electronic analog computer which utilized electronic function generators to modify the transfer function of the electronic signals passing through wires. The high density monolithic embedded optical fiber interconnect method makes this approach very attractive since noise and crosstalk are eliminated. A reference light source transmitted through optical waveguides can be altered with "optical transfer function generators" as they pass through recursive loops in this high density, high speed media. The function generators modulate the optical signal parameters to produce any transfer function which is the mathematical relationship between the signal input and signal output. As is the case with conventional analog computers, a combination of amplification, attenuation, energy storage and signal splitting can be used to generate or alter the desired linear or non-linear functions. Optical amplifiers are currently being developed for optical communication providing useful gain in the optical domain without conversion of signals to electronic format. Devices are also becoming available which make possible the storage of optical energy such as fiber fabry-perot interferometers and/or optical Q switches. Until these all optical devices are available in suitable format, resistive ladders, variable resistors, switched capacitor banks, or tapped inductors can also be utilized to control optical function generators which can be electronically modified in real time. Thus the amplitude, phase, and frequency of the transfer function can be optically or electronically controlled. Multiple transfer functions can be serially cascaded or activated in parallel to provide real time signal processing or computation. This method obviates many noise, impedance, and accuracy limitations of previous electronic analog computers by eliminating patchcords, electronic patchcords, and electronic "tweaking" since the information signal is in the optical domain. Optical signals can also be utilized to directly vary the resistance, capacitance, or inductance of the electronic control circuit elements. 4.3 Modular Electro-Optic Packaging A modular packaging concept for the electro-optic processors and control elements described above can be implemented in a method compatible with embedded optical fiber backplanes. As described previously, the optical backplane could be totally passive and contain only the reconfigurable waveguides, the fanout couplers and high density connectors. These backplanes would then interface with a module containing the laser/led arrays and the photodetectors to provide the dynamic interconnect. Figure 15 shows such a module. Power and ground connections are shown as connecting to the backplane but they could also be made to interface from the sides or top of the module to keep the optical backplane totally passive if desired. 5.0 CONCLUSIONS AND RECOMMENDATIONS Over the years, the human brain has been compared to the technological developments of the time including the telephone switchboard and the digital computer. More recently this analogy has been extended to complex information transfer networks. I have attempted to extend this analogy to optical waveguide networks and the comparable elements are shown in figure 16. Optical technology, specifically embedded optical fiber backplanes, can provide enormous cost, reliability, and performance advantages over the multilayer wiring boards currently used in high speed computers. The real time topological reconfiguration of these backplanes through switched laser/led and photodetector arrays provides a capability not achievable with other interconnect technologies. The rapid development of this interconnect technology will provide high payoff in parallel connectionist approaches to computing and signal processing including neural networks. Optical activation and reconfiguration of processors utilizing high density VLSI digital and analog microcircuit technology can add another dimension of adaptive control to connectionist parallel processing. This dimension can be used to reduce the software complexity of existing machines through the direct functional mechanization of recursive transfer networks. As optical component technology matures, one can envision an evolutionary transition from optical interconnects for electronic processors to an all optical digital, analog, and hybrid computing capability. A major development initiative can hasten this transition. 6.REFERENCES 1. R. M. Restak, The Brain. The Last Frontier, Chapters 7-9, Warner books, 1980. 2. R. Ornstein and R. F. Thompson, The Amazing Brain, Chapters 3-4, Houghton Mifflin Company, 1984. 3. A. S. Glista, "Fault Tolerant Topologies for Fiber Optic Networks and Computer Interconnects Operating in the Severe Avionics Environment", IEEE Journal of Lightwave Communications Systems, pp.66-78, Feb. 1991 Figure 1. Figure 2 Figure 3. Figure 4. Figure 5. figure 6. Figure 7. Figure 8. Figure 9. Figure 10. Figure 11. Figure 12 Figure 13. Figure 14. Figure 15. Figure 16.