Granted Patent B2

US 12,353,988 · App. 17/305,486 · Granted Jul 8, 2025

Inventors: Nikolaos Pleros (Salonika, GR); George Mourgias-Alexandris (Salonika, GR); Angelina Totovic (Central Macedonia, GR); Apostolos Tsakyridis (Salonika, GR); Nikolaos Passalis (Salonika, GR); Anastasios Tefas (Salonika, GR); Konstantinos Vyrsokinos (Salonika, GR)

Assignee: Celestial AI Inc.

G06N3/0675H04B10/516H04B10/548G06N3/049H04B10/54

Abstract

Claims (44)

1. An optical neuron to optically process a plurality of neuron inputs, the optical neuron comprising:

a multipath optical interferometer comprising:

an optical splitter;

a plurality of parallel optical interferometer branches coupled to the optical splitter, the optical splitter configured to operatively split light between the parallel optical interferometer branches into a plurality of optical carrier signals, the plurality of parallel optical interferometer branches comprising multiple input branches each associated with one of the neuron inputs and each comprising a first amplitude modulator, a second amplitude modulator, and at least one phase shifter; and

an optical combiner coupled to outputs of the plurality of parallel optical interferometer branches and configured to operatively cause interference between modulated optical carrier signals received from the plurality of parallel optical interferometer branches to generate an optical interference signal; and

electronic driver circuitry configured to operatively control the first amplitude modulators, the second amplitude modulators, and the phase shifters to:

cause the first amplitude modulators to impart the neuron inputs onto field amplitudes of the optical carrier signals in the associated input branches,

cause the second amplitude modulators to impart neuron weights onto the field amplitudes of the optical carrier signals in the associated input branches, and

cause the phase shifters to selectively induce zero or n phase shifts in the optical carrier signals in the associated input branches to thereby impart signs of the neuron weights and neuron inputs onto the optical carrier signals in the associated input branches,

wherein the optical interference signal is indicative of a weighted sum of the neuron inputs.

2. The optical neuron of claim 1 , wherein the plurality of parallel optical interferometer branches further comprises a bias branch comprising an amplitude modulator to impart a weighted bias onto a field amplitude of the optical carrier signal in the bias branch, the weighted bias operatively set to cause a field amplitude of the optical carrier signal in the bias branch to exceed a maximum expected absolute value of a negative coherent sum of the field amplitudes of the modulated optical carrier signals output by the input branches.

3. The optical neuron of claim 1 , further comprising:

a non-linear activation and conversion unit coupled to an output of the optical combiner and comprising a photodetector to generate an electronic neuron output signal from the optical interference signal.

4. The optical neuron of claim 3 , wherein the non-linear activation and conversion unit comprises an all-optical activation unit preceding the photodetector.

5. The optical neuron of claim 3 , wherein the non-linear activation and conversion unit comprises an electronic activation unit following the photodetector.

6. The optical neuron of claim 1 , wherein the first and second amplitude modulators comprise at least one of a Mach-Zehnder modulator, an electro-absorption modulator, or an optical resonant modulator.

7. The optical neuron of claim 1 , wherein the phase shifters comprise at least one of a thermal phase shifter or an electro-optic phase shifter.

8. The optical neuron of claim 1 , wherein the multipath interferometer is implemented as part of an integrated photonic circuit.

9. An optical neural network comprising:

a plurality of optical neurons each comprising:

a multipath interferometer comprising a plurality of parallel optical interferometer branches coupled between an optical splitter and an optical combiner, the optical interferometer branches comprising a plurality of input branches associated with neuron inputs and each comprising a first amplitude modulator, a second amplitude modulator, and at least one phase shifter, the optical combiner configured to operatively generate an optical interference signal from modulated optical carrier signals output by the plurality of optical interferometer branches;

electronic driver circuitry configured to operatively control the first and second amplitude modulators and the phase shifters in the input branches to impart the signed neuron inputs and signed neuron weights onto optical carrier signals in the respective input branches to generate the modulated optical carrier signals output by the input branches; and

a non-linear activation and conversion unit, coupled to an output of the optical combiner and comprising a photodetector, to generate an electronic neuron output signal; and

electronic control circuitry configured to determine, from the electronic neuron output signal of at least one of the plurality of optical neurons, a neuron input to another one of the plurality of optical neurons.

10. The optical neural network of claim 9 , wherein the electronic driver circuitry of each of the optical neurons is configured to operatively:

cause the first amplitude modulators to impart the neuron inputs onto field amplitudes of the optical carrier signals in the associated input branches,

cause the second amplitude modulators to impart values of the signed neuron weights onto the field amplitudes of the optical carrier signals in the associated input branches, and

cause the phase shifters to selectively induce zero or 7 C phase shifts in the optical carrier signals in the associated input branches to thereby impart signs of the signed neuron weights and the signed neuron inputs onto the optical carrier signals in the input branches.

11. The optical neural network of claim 9 , wherein, for each of the optical neurons, the optical interferometer branches further comprise a bias branch comprising an amplitude modulator to operatively impart a weighted bias onto an optical carrier signal in the bias branch to generate the modulated optical carrier signal output by the bias branch, the weighted bias operatively set to cause a field amplitude of the modulated optical carrier signal in the bias branch to exceed a maximum expected absolute value of a negative coherent sum of field amplitudes of the modulated optical carrier signals output by the input branches.

12. The optical neural network of claim 9 , wherein at least one of the non-linear activation and conversion units of the plurality of optical neurons comprises an all-optical activation unit generating an optical neuron output signal preceding the photodetector.

13. The optical neural network of claim 9 , wherein the optical neurons are implemented with integrated photonic circuitry on a single substrate.

14. The optical neural network of claim 13 , wherein the electronic control circuitry is implemented as an integrated circuitry packaged with the integrated photonic circuitry in a system-on-chip or chip-scale package.

15. An optical neuromorphic computing method employing an optical neuron comprising a multipath optical interferometer, the method comprising:

splitting light between multiple optical interferometer branches of the multipath optical interferometer into multiple respective optical carrier signals, each of the multiple optical interferometer branches having an associated neuron input and an associated neuron weight;

modulating field amplitudes and phases of the optical carrier signals in the optical interferometer branches to separately impart the associated neuron inputs and the associated neuron weights onto the optical carrier signals; and

recombining the optical carrier signals at outputs of the optical interferometer branches to generate an optical interference signal indicative of a weighted sum of the neuron inputs.

16. The optical neuromorphic computing method of claim 15 , wherein the field amplitudes of the optical carrier signals are modulated by first amplitude modulators in the optical interferometer branches to impart the associated neuron inputs and by second amplitude modulators in the optical interferometer branches to separately impart the associated neuron weights onto the optical carrier signals, and wherein the phases of the optical carrier signals in the input are selectively shifted by zero or π to thereby impart signs of the associated neuron inputs and neuron weights onto the optical carrier signals.

17. The optical neuromorphic computing method of claim 16 , wherein the field amplitudes and phases of the optical carrier signals are modulated at least one of electro-optically or thermo-optically.

18. The optical neuromorphic computing method of claim 15 , further comprising:

measuring an intensity of the optical interference signal or of an optical neuron output signal generated from the optical interference signal, and using the measured intensity to determine a neuron input to another optical neuron.

19. The optical neuromorphic computing method of claim 18 , further comprising:

applying a non-linear optical activation function to the optical interference signal to generate the optical neuron output signal.

20. The optical neuromorphic computing method of claim 15 , further comprising:

biasing the optical interference signal to encode both absolute value and sign of the weighted sum of the neuron inputs in an intensity of the optical interference signal.

Assignments (3)

Continuity (2)

References Cited (344)