Apparatus and methods for optical neural network
An optical neural network is constructed based on photonic integrated circuits to perform neuromorphic computing. In the optical neural network, matrix multiplication is implemented using one or more optical interference units, which can apply an arbitrary weighting matrix multiplication to an array of input optical signals. Nonlinear activation is realized by an optical nonlinearity unit, which can be based on nonlinear optical effects, such as saturable absorption. These calculations are implemented optically, thereby resulting in high calculation speeds and low power consumption in the optical neural network.
1. An apparatus for implementing an artificial neural network, the apparatus comprising:
an array of input waveguides to receive a first array of optical signals;
an optical interference unit, in optical communication with the array of input waveguides, to perform a matrix multiplication of the first array of optical signals with a transformation matrix to produce a second array of optical signals;
an array of output waveguides, in optical communication with the optical interference unit, to guide the second array of optical signals, wherein at least one input waveguide in the array of input waveguides is in optical communication with each output waveguide in the array of output waveguides via the optical interference unit; and
control circuitry to adjust weight parameters of the transformation matrix implemented by the optical interference unit.
2. The apparatus of claim 1 , wherein the optical interference unit comprises:
a plurality of interconnected Mach-Zehnder interferometers (MZIs), each MZI in the plurality of interconnected MZIs comprising:
a first phase shifter configured to change a splitting ratio of the MZI; and
a second phase shifter configured to shift a phase of one output of the MZI.
3. The apparatus of claim 2 , wherein the plurality of interconnected MZIs is configured to perform the matrix multiplication of the first array of optical signals via singular value decomposition (SVD).
4. The apparatus of claim 2 , wherein the plurality of interconnected MZIs comprises:
a first set of MZIs to perform a unitary transformation of the first array of optical signals to generate a first array of transformed optical signals; and
a second set of MZIs in optical communication with the first set of MZIs, each MZI in the second set of MZIs receiving a corresponding transformed optical signal and transmitting a first output to a nonlinearity unit,
wherein a second output of each MZI in the second set of MZIs is optically blocked so as to change a total intensity of the corresponding transformed optical signals and perform a non-unitary diagonal matrix multiplication of the corresponding transformed optical signals.
5. The apparatus of claim 1 , wherein the optical interference unit comprises:
a plurality of interconnected Mach-Zehnder interferometers (MZIs) to perform a unitary transformation of the first array of optical signals; and
an array of optical attenuators or amplifiers, in optical communication with the plurality of interconnected MZIs, to change an intensity of the first array of optical signals after the plurality of interconnected MZIs.
6. The apparatus of claim 1 , further comprising:
an optical nonlinearity unit, in optical communication with the optical interference unit, to perform a nonlinear transformation on the second array of optical signals so as to generate a third array of optical signals.
7. A method for artificial neural network computation, the method comprising:
receiving a first array of optical signals with an array of input waveguides;
interfering the first array of optical signals, using an optical interference unit in optical communication with the array of input waveguides, to linearly transform the first array of optical signals into a second array of optical signals;
guiding the second array of optical signals using an array of output waveguides, wherein at least one input waveguide in the array of input waveguides is in optical communication with each output waveguide in the array of output waveguides via the optical interference unit; and
using control circuitry to adjust weight parameters of a transformation matrix implemented by the optical interference unit to linearly transform the first array of optical signals into the second array of optical signals.
8. The method of claim 7 , wherein interfering the first array of optical signals comprises:
propagating the first array of optical signals through a plurality of interconnected Mach-Zehnder interferometers (MZIs);
changing a splitting ratio of at least one MZI in the plurality of interconnected MZIs; and
shifting a phase of one output of the at least one MZI.
9. The method of claim 8 , wherein propagating the first array of optical signals through the plurality of interconnected MZIs comprises performing a linear transformation on the first array of optical signals via singular value decomposition (SVD).
10. A method for artificial neural network computation, the method comprising:
receiving a first array of optical signals with an array of input waveguides;
interfering the first array of optical signals, using an optical interference unit in optical communication with the array of input waveguides, to linearly transform the first array of optical signals into a second array of optical signals;
guiding the second array of optical signals using an array of output waveguides, wherein at least one input waveguide in the array of input waveguides is in optical communication with each output waveguide in the array of output waveguides via the optical interference unit; and
implementing a nonlinear transformation of the second array of optical signals in the electronic domain.
11. The method of claim 10 , wherein implementing a nonlinear transformation of the second array of optical signals in the electronic domain comprises measuring optical mode output intensities on a photodetector array and injecting signals into the next stage.
12. The method of claim 11 , further comprising:
using control circuitry to adjust weight parameters of a transformation matrix implemented by the optical interference unit to linearly transform the first array of optical signals into a second array of optical signals.
13. The method of claim 11 , wherein interfering the first array of optical signals comprises:
propagating the first array of optical signals through a plurality of interconnected Mach-Zehnder interferometers (MZIs);
changing a splitting ratio of at least one MZI in the plurality of interconnected MZIs; and
shifting a phase of one output of the at least one MZI.
14. The method of claim 13 , wherein propagating the first array of optical signals through the plurality of interconnected MZIs comprises performing a linear transformation on the first array of optical signals via singular value decomposition (SVD).
15. The method of claim 11 , wherein interfering the first array of optical signals comprises:
performing a unitary transformation of the first array of optical signals with a plurality of interconnected MZIs; and
attenuating or amplifying the first array of optical signals with an array of optical attenuators in optical communication with the plurality of interconnected MZIs.