The disclosed circuit and method relate to optical circuits. More specifically, the disclosed circuit and method relate to an all-optical circuit for performing computations.
Signal processing continues to become more complex as data transfer rates continue to increase. Conventionally, signal processing is either performed by an analog system or by a digital system. Analog systems may be implemented in compact circuits making them a popular choice when circuit area is of significant importance. However, one significant drawback of analog devices is that they are highly susceptible to noise accumulation, which limits the number of analog operations that can be applied to data, and therefore the complexity of the computations that can be practically implemented using only analog devices.
In contrast, digital systems are not as susceptible to noise accumulation as are analog systems. However, the number of digital devices needed to implement a computation rapidly increases with the complexity of the computation performed.
Photonic devices provide the ability to process signals of much higher bandwidth than is possible with electronic devices, but they are larger and more expensive than electronic devices. Practical implementation of complex high bandwidth processing algorithms using photonic devices therefore requires an approach that minimizes the number of devices needed without over constraining the complexity of computations that can be implemented.
Accordingly, a hybrid processing system and method that combines the advantages of digital and analog systems is desirable.
An optical system is disclosed that includes an optical integrator, a readout mechanism, and an optical thresholder. The optical integrator is configured to perform temporal integration of an optical input signal having a first wavelength received at an input. The readout mechanism is coupled to the optical integrator and provides optical signals having a second wavelength to the optical integrator for measuring a state of the optical integrator. The optical thresholder is coupled to an output of the optical integrator and is configured to receive a signal representing a temporal integration of the optical input signal from the optical integrator and produce an optical signal identifying if an amplitude of the signal representing the temporal integration of the optical input signal is above or below a threshold value.
A signal processing method is also disclosed. The optical signal processing method includes temporally integrating a first optical signal having a first wavelength at an optical integrator, determining if the temporally integrated optical signal has an amplitude that is above a threshold at an optical thresholder, and outputting an optical signal identifying if the amplitude of the temporally integrated optical signal is above or below the threshold.
Additionally, an optical system is disclosed including a semiconductor optical amplifier (SOA), an optical filter, and an optical thresholder. The SOA has a decaying response function for temporally integrating an optical signal having a first wavelength received at an input. The SOA is configured to output an optical signal identifying a state of the SOA in response to receiving a signal having a second wavelength from a readout device. The optical filter is coupled to an output of the SOA and is configured to pass the optical signals having the second wavelength and blocking optical signals having the first wavelength. The optical thresholder is configured to receive optical signals having the second wavelength identifying the state of the SOA from the optical filter and provide an optical signal identifying if an energy of the optical signals having the second wavelength are above or below a threshold value.
The leaky integrate-and-fire (“LIF”) neuron is one of the most widely studied neuron models in computational neuroscience. The spike processing performed by these computational elements is a hybrid of analog and digital processing that exploits the efficiency of analog computation while overcoming the problem of noise accumulation suffered by analog systems. In spike processing, information is encoded in the timing of spikes rather than in the size or shape of the spikes. Accordingly, information is conveyed by a spike being present or absent, much like in digital systems how a bit is either a one or a zero. In contrast, traditional neural network models perform purely analog computation in which neuron inputs, intermediate results, and outputs are represented as analog values.
From the standpoint of computability and complexity theory, LIF neurons are powerful computational primitives capable of simulating both Turing machines and traditional neural networks. LIF models have a number, N, of inputs, σi(t), where i=1,2, . . . , N; an internal activation state, Vm(t); and a single output state, O(t). At rest, the internal state of the neuron is actively maintained at a resting voltage, Vrest. Each input, σt(t), of the neuron is a continuous time series consisting of either spikes or continuous analog values. These inputs are typically weighted by ωi and delayed δi, which may be mathematically represented as ωiσi(t+δi). The delayed and weighted input time series is spatially integrated through pointwise summation in accordance with the following equation:
Σi=1Nωiσi(t+δi) Eq. (1)
The activation state, or membrane voltage, Vm(t), of the neuron is an exponentially weighted temporal integration of the spatially summed input time series. If the magnitude of the temporally integrated signal exceeds a threshold value, then the neuron outputs a spike, e.g., O(t)=1, if Vm(t)>Vthresh. After a spike, there is a short period of time, known as a refractory period, during which another spike cannot be issued, e.g., if O(t)=1, then O(t+Δt)=0, Δt≦Trefract. Accordingly, the output of the neuron, O(t), consists of a continuous time series of spikes.
There are three primary influences that affect the magnitude of the membrane voltage, Vm(t), of an LIF neuron: (1) an active pumping current, (2) current leakage, and (3) external inputs generating time varying membrane conductance changes. Each of these influences are part of the following differential equation for approximating the membrane voltage over time:
represents the activation of the neuron;
represents the active pumping current of the neuron;
represents the leakage current of neuron; and
represents the external inputs to the neuron.
A direct correspondence has been discovered between the equation governing temporal integration of LIF neurons set forth in Equation 2 and the carrier density of a semiconductor optical amplifier (SOA). The primary state variable for a SOA in this case is the carrier density above transparency N′(t)=N(t)−N0, where N(t) is the actual carrier density, and N0(t) is carrier density at transparency. The integrative properties of the SOA are determined by the carrier lifetime, τe; a mode confinement factor, ┌; a differential gain coefficient, α; a photon energy, Ep; and the active SOA pumping current, I(t). Spontaneous carrier decay tends to drive the carrier density, N′, of the SOA towards zero, and thus the active pumping current is needed to counter the carrier decay to maintain a resting carrier density of N′rest. The three contributions to the value of the carrier density, N′(t), are a leakage term due to passive carrier decay, a term for carrier density due to active optical pumping of the SOA, and a term for carriers generated by external inputs. The gain dynamics of a SOA when input pulse widths are much shorter, e.g., two orders of magnitude shorter, than the carrier lifetime may be described as follows:
Comparing Equation 2 with Equation 3 demonstrates a remarkable similarity between the electrical model of membrane voltage of an LIF neuron and the optical model of SOA carrier density. The discovery of the correspondence between the Equations 1 and 2 has enabled the development of an all-optical implementation of an LIF neuron, which can advantageously be used as a computational primitive in large scale complex photonic computational systems.
The thresholder 110 includes a non-linear optical loop mirror 122 formed by a non-linear doped fiber 124 and an optical coupler 126. A plurality of optical components may be disposed along non-linear fiber 124 for optical tuning. For example, polarization controllers 128-1, 128-2 and a tunable isolator 130 may be disposed along the non-linear fiber 124.
Optical couplers 106 (e.g., optical couplers 106-1 and 106-2) may be any optical coupler configured to couple optical signals of different wavelengths and amplitudes in separate fibers into a single fiber. In one example, the optical coupler 106-1 is an N:1 optical coupler configured to couple N optical input signals into a single fiber, and optical coupler 106-2 is an optical coupler configured to couple the optical input signals output from optical coupler 106-1 with optical signals from a pulse train provided by the optical readout mechanism 112 into a single fiber. An example of a suitable fiber coupler 106-2 is a thermally tapered and fused pair of single-mode fibers, with the cores of the fiber pair coming into contact such that optical energy may be exchanged. Multiport coupler 106-1 may be, for example, a tree of 2:1 couplers as will be understood by one skilled in the art. The optical signals of the sampling pulse train may have a wavelength λ0, and the optical input signals may have one or more wavelengths λ1, λ2, etc., which are different from the wavelength of the pulse train. Additionally, the optical input signals have amplitudes that are greater than the amplitudes of the optical signals of the pulse train such that the optical signals of the pulse train do not have a significant effect on the cross-gain modulation (XGM) of the SOA 114 as described below. For example, the amplitudes of the optical input signals may be ten times the amplitude of the optical signals of the pulse train. However, one skilled in the art will understand that the difference between the amplitudes of the optical input pulses and the optical signals of the pulse train may be increased or decreased.
Readout mechanism 112 (
One example of an SOA 114 is illustrated in
Referring again to
The optical amplifier 120 may be any optical amplifier configured to increase the amplitude of an optical signal. In one arrangement, the optical amplifier 120 is an erbium-doped fiber amplifier (EDFA). Another example of optical amplifier 120 is an erbium-ytterbium-doped fiber amplifier, which may provide a higher output power than an EDFA.
The non-linear fiber 124 of thresholder 110 may be a GeO2-doped silica-based non-linear fiber. Other non-linear fibers such as, for example, microstructured fibers, photonic crystal fibers, and hi-oxide fibers, to name a few, may also be implemented. Parasitic reflections may be suppressed for proper thresholder operation as will be understood by one skilled in the art.
Polarization controllers 128-1, 128-2 may use a controllable bend of fiber to control the polarization of light and may be based on a medium having a weak controllable birefringence. Tunable isolator 130 may be any kind of optical isolator and may be based on the Faraday effect and have a controllable leak in a backward direction. Coupler 126 may be the same type of optical coupler as couplers 106-1, 106-2 except that it may have an unequal coupling ratio. Examples of coupling ratios include, but are not limited to, an 80:20 ratio, a 90:10 ratio, to name a few.
An optional optical inverter 132 may be coupled to the output of the thresholder 110 for inverting the thresholder output. In some embodiments, the optical inverter 132 may be an optical logic gate such as the one taught by Miyoshi et al. in Ultrafast All-Optical Logic Gate Using a Nonlinear Optical Loop Mirror Based Multi-Periodic Transfer Function, Optics Express, Vol. 16, Issue 4, 2570-2577 (2008), the entirety of which is incorporated by reference herein. The optical logic gate may be configured to output an optical signal having the same characteristics of the optical input signals (e.g., wavelength and amplitude) in response to the output of the thresholder 110. One skilled in the art will understand that other optical devices may be used to invert the optical signal output from the thresholder 110. For example, the optical inverter 132 may include an SOA for inverting the optical signal or an optical data format converter based on a tetrahertz optical asymmetrical demultiplexer (TOAD) such as the one disclosed in U.S. Pat. No. 6,448,913 issued to Prucnal at al., the entirety of which is incorporated by reference herein.
With reference to
Optical coupler 106-2 outputs a multiple wavelength optical signal to an input of the SOA 114. SOA 114 of the integration block 108 is pumped with electrons from a charge pump circuit 118, which performs population inversion of the SOA 114. When a pulse from one of the optical input signals is received at the SOA 114, the gain of the SOA 114 is depleted due to the depletion of charge that occurs due to XGM. The external pumping of the SOA 114 causes the gain of the SOA 114 to gradually increase, but if another pulse is received from the optical input signals, then the gain of the SOA 114 will again be depleted. The recovery time of the gain of the SOA 114 is based on the carrier lifetime, Te, which functions as the integration time constant of the integrator 108. Thus, the smaller the carrier lifetime of the SOA 114 the faster the gain of the SOA 114 recovers and less temporal integration of the input signals is performed.
The integrated optical signal is output from the SOA 114 and received at the optical filter 116. As described above, the optical filter 116 may be tuned such that the optical filter 116 passes the optical signals having a wavelength λ0 and the optical input signals are blocked, reflected, or otherwise filtered out.
The filtered optical signal is received at an input of the optical amplifier 120, which may be an EDFA as described above. The optical amplifier increases the amplitude of the filtered and integrated optical signal and outputs the amplified optical signal to the thresholder 110. The gain of the optical amplifier 120 may be designed to provide sufficient amplification of the filtered optical signal so that the amplitude of the signal falls within the linear region of the non-linear fiber 124 of the thresholder 110.
Thresholder 110 outputs a signal identifying if the optical signal received at the input is greater than or equal to a predetermined threshold level. However, due to the utilization of the SOA 114 with gain sampling, the output of the thresholder when one of the optical input signals is a logic one, the thresholder will output a logic zero.
As described above, an optical inverter 132 may be coupled to the output of the thresholder 110 to invert the signal output from the thresholder 110. For example, if the inverter 132 is an optical logic gate such as the ones described by Miyoshi et al., the inverter may output an optical signal having a wavelength of λ1 and an amplitude equal to the amplitude of the optical input signal when the thresholder 110 outputs a logic zero. The output of the optical inverter 132 may be fed into another processing element as will be understood by one skilled in the art. For example, the output of optical inverter 132 may be fed into another thresholder to provide standardization of the pulses at the neuron output according to the LIF neuron model.
An optical LIF neuron 100 as illustrated in
The gain dynamics of the tested SOA 114 are shown in
The output of the integrator 108 is illustrated in
The thresholder 110 was constructed using a non-linear fiber based on a modified nonlinear optical loop mirror. The thresholder 110 included 10.5 m of a silica-based non-linear fiber 124 heavily-doped with GeO2 (preform 311). The fiber parameters measured at λ=1550 nm were a nonlinear coefficient 35 W−1 km−1; propagation losses of 36 dB/km; chromatic dispersion −70 ps/nm km; and a refraction index difference, Δn, of approximately 0.11. The idealized model of the thresholder 110 predicted that the output power was proportional to the cube of the input power, A measured transfer function had a cubic dependence for some range of input powers with saturation at a certain input level at which the nonlinear phase shift approaches π. The input power of the thresholder was controlled by adjusting the gain of the optical amplifier 120, which was implemented as an EDFA with a maximum output power of approximately 23 dBm.
The neuron 100 was able to discriminate between the lowest and middle pulse energies, e.g., a zero or a one, or between the middle and the highest pulse energies, e.g., a one and a two. Both possibilities were experimentally demonstrated in the setup with corresponding diagrams shown in
The experimentally demonstrated photonic LIF device was shown to be operable using picosecond-width pulses and have an integration time constant of 180 ps, which was adjustable within the range of approximately 100 to 300 ps. Reconfiguration of device parameters enables it to perform a wide variety of signal processing and decision operations, its analog properties makes it well-suited for efficient signal processing applications. The digital properties of the optical LIF neuron make it possible to implement complex computations without excessive noise accumulation.
Although the systems and methods have been described in terms of exemplary embodiments, they are not limited thereto. Rather, the appended claims should be construed broadly, to include other variants and embodiments of the systems and methods, which may be made by those skilled in the art without departing from the scope and range of equivalents of the systems and methods.
This application claims priority to U.S. Provisional Patent Application No. 61/158,986, which was filed on Mar. 10, 2009, and is incorporated by reference herein in its entirety.
Filing Document | Filing Date | Country | Kind | 371c Date |
---|---|---|---|---|
PCT/US10/26835 | 3/10/2010 | WO | 00 | 10/26/2011 |
Number | Date | Country | |
---|---|---|---|
61158986 | Mar 2009 | US |