Radio-frequency photonic architecture for deep neural networks, signal processing, and computing
A multiplicative analog frequency transform optical neural network (MAFT-ONN) encodes data in the frequency domain, achieves matrix-vector products in a single shot using photoelectric multiplication, and uses a single electro-optic modulator for the nonlinear activation of all neurons in each layer. Photoelectric multiplication between radio frequency (RF)-encoded optical frequency combs allows single-shot matrix-vector multiplication and nonlinear activation, leading to high throughput and ultra-low latency. This frequency-encoding scheme can be implemented with several neurons per hardware spatial mode and allows for an arbitrary number of layers to be cascaded in the analog domain. For example, a three-layer DNN can compute over four million fully analog operations and implement both a convolutional and fully connected layer. Additionally, a MAFT-ONN can perform analog DNN inference of temporal waveforms like voice or radio signals, achieving bandwidth-limited throughput, speed of light-limited latency, and fully analog complex-valued matrix operations.
1 . A method of performing a computation on an input vector and a matrix, the method comprising:
frequency encoding the input vector onto a first optical signal;
frequency encoding the matrix onto a second optical signal; and
detecting a heterodyne interference signal between the first optical signal and the second optical signal, the heterodyne interference signal including frequency-encoded products of elements of the input vector and elements of the matrix and spurious frequency content.
2 . The method of claim 1 , wherein the computation is a multiplication of the input vector and the matrix, the heterodyne interference signal further includes spurious frequency content, and further comprising:
filtering the spurious frequency content from the heterodyne interference signal.
3 . The method of claim 2 , wherein filtering the spurious frequency content from the heterodyne interference signal comprises bandpass filtering the heterodyne interference signal.
4 . The method of claim 2 , wherein filtering the spurious frequency content from the heterodyne interference signal comprises applying periodic filtering to the heterodyne interference signal.
5 . The method of claim 1 , wherein the computation is a convolution of the input vector and the matrix.
6 . The method of claim 1 , wherein frequency encoding the input vector onto the first optical signal comprises modulating the first optical signal with a single-sideband suppressed-carrier (SSB-SC) modulator driven by a frequency-encoded version of the input vector.
7 . The method of claim 6 , wherein the first SSB-SC modulator is a dual-parallel Mach-Zehnder modulator comprising a first Mach-Zehnder modulator (MZM) in a first arm of a Mach-Zehnder interferometer and a second MZM in a second arm of the Mach-Zehnder interferometer and modulating the first optical signal with the first SSB-SC modulator comprises:
driving the first MZM with an in-phase component of the frequency-encoded version of the input vector; and
driving the second MZM with a quadrature component of the frequency-encoded version of the input vector.
8 . The method of claim 1 , further comprising:
selecting a frequency encoding of the matrix to yield the frequency-encoded products at frequencies interspersed with the spurious frequency content.
9 . The method of claim 1 , further comprising:
selecting a frequency encoding of the matrix to yield the frequency-encoded products at frequencies in a band that does not include any of the spurious frequency content.
10 . The method of claim 1 , wherein the input vector is an input activation vector to a layer of a neural network, the matrix is a weight matrix of the neural network, the frequency-encoded products represent an output of the layer of the neural network, and further comprising:
modulating a third optical signal with the frequency-encoded products as an input activation vector of a subsequent layer of the neural network.
11 . The method of claim 10 , wherein modulating the third optical signal comprises driving a Mach-Zehnder modulator with the frequency-encoded products, and further comprising:
applying a nonlinearity in the subsequent layer of the neural network via the Mach-Zehnder modulator.
12 . An optical processor comprising:
a laser to emit an optical carrier;
a first single-sideband suppressed carrier (SSB-SC) modulator, in optical communication with the laser, to modulate a first copy of the optical carrier with a frequency-encoded version of an input vector as a first optical signal;
a second SSB-SC modulator, in optical communication with the laser, to modulate a second copy of the optical carrier with a frequency-encoded version of a matrix as a second optical signal;
a beam splitter, in optical communication with the first SSB-SC modulator and the second SSB-SC modulator, to combine the first optical signal and the second optical signal; and
a photodetector, in optical communication with the beam splitter, to detect a heterodyne interference signal between the first optical signal and the second optical signal, the heterodyne interference signal including frequency-encoded products of elements of the input vector and elements of the matrix.
13 . The optical processor of claim 12 , wherein the first SSB-SC modulator is a dual-parallel Mach-Zehnder modulator (DPMZM) that comprises a first Mach-Zehnder modulator (MZM) in a first arm of a Mach-Zehnder interferometer and driven by an in-phase component of the frequency-encoded version of the input vector and a second MZM in a second arm of the Mach-Zehnder interferometer and driven by a quadrature component of the frequency-encoded version of the input vector.
14 . The optical processor of claim 12 , wherein the photodetector is a balanced differential photodetector.
15 . The optical processor of claim 12 , wherein the heterodyne interference signal further includes spurious frequency content, and further comprising:
a filter, operably coupled to the photodetector, to filter the spurious frequency content from the heterodyne interference signal.
16 . The optical processor of claim 12 , wherein the input vector is an input to a first layer of a neural network, the matrix is a weight matrix of the neural network, and further comprising:
a third SSB-SC modulator, operably coupled to the photodetector, to modulate a third copy of the optical carrier with heterodyne interference signal as an input to a second layer of the neural network.
17 . The optical processor of claim 16 , wherein the third SSB-SC modulator is configured to apply a nonlinearity of the neural network.
18 . The optical processor of claim 12 , wherein the laser is a first laser, the optical carrier is a first optical carrier at a first wavelength, the input vector is a first input vector, the matrix is a first matrix, and the heterodyne interference signal is a first heterodyne interference signal, and further comprising;
a second laser to emit a second optical carrier at a second wavelength different that the first wavelength;
a third SSB-SC modulator, in optical communication with the second laser, to modulate a first copy of the second optical carrier with a frequency-encoded version of a second input vector as a third optical signal;
a fourth SSB-SC modulator, in optical communication with the second laser, to modulate a second copy of the second optical carrier with a frequency-encoded version of a second matrix as a fourth optical signal;
a first multiplexer, operably coupled to the first SSB-SC modulator and the third SSB-SC modulator, to multiplex the first optical signal and the third optical signal onto the beam splitter; and
a second multiplexer, operably coupled to the second SSB-SC modulator and the fourth SSB-SC modulator, to multiplex the second optical signal and the fourth optical signal onto the beam splitter,
wherein the photodetector is configured to detect a second heterodyne interference signal between the third optical signal and the fourth optical signal, the second heterodyne interference signal including frequency-encoded products of elements of the second input vector and elements of the second matrix.
19 . The optical processor of claim 12 , wherein the input vector is a first input vector, the beam splitter is a first beam splitter, the photodetector is a first photodetector, the heterodyne interference signal is a first heterodyne interference signal, and further comprising;
a third SSB-SC modulator, in optical communication with the laser, to modulate a third copy of the optical carrier with a frequency-encoded version of a second input vector as a third optical signal;
a second beam splitter, in optical communication with the third SSB-SC modulator and the second SSB-SC modulator, to combine the third optical signal and the second optical signal; and
a second photodetector, in optical communication with the second beam splitter, to detect a second heterodyne interference signal between the third optical signal and the second optical signal, the heterodyne interference signal including frequency-encoded products of elements of the second input vector and elements of the matrix.
20 . The optical processor of claim 12 , further comprising:
a switch, operably coupled to an output of the photodetector and an input to the first SSB-SC modulator, to switch the heterodyne interference signal between the input to the first SSB-SC modulator and an output of the optical processor.