Optical diffractive processing unit

1. A optical diffractive processing unit, comprising:

input nodes;

output nodes;

neurons, connected to the input nodes through optical diffractions, wherein weights of connection strength of the neurons are determined based on diffractive modulation, each optoelectronic neuron is configured to perform an optical field summation of weighted inputs and generate a unit output by applying a complex activation to an optical field occurring naturally in a photoelectronic conversion, and the neurons are formed by a programmable device;

the programmable device comprises an optical neural network containing a digital micromirror device, a spatial light modulator, and a photodetector;

the digital micromirror device is configured to provide a high optical contrast for information coding;

the spatial light modulator is configured to perform the diffractive modulation, wherein weighted connections between the input nodes and the neurons are implemented by free-space optical diffraction, and a receiving field of each neuron is determined by an amount of diffraction from a plane of the spatial light modulator to a plane of the photodetector; and

the photodetector is configured to implement the optical field summation and the complex activation.

2. The optical diffractive processing unit of claim 1 , wherein the optical neural network comprises multiple single-layer diffractive layers, and each single-layer diffractive layer comprises an input coding layer, a diffractive connection layer, an optical summation layer, and an optical non-linearity layer.

3. The optical diffractive processing unit of claim 2 , wherein the input coding layer is implemented by a programmable input module configured to encode input data into incident light, a physical dimension of encoding comprises amplitude encoding, phase encoding or both the amplitude encoding and the phase encoding, and an encoding type comprises discrete encoding, continuous encoding, or both the discrete encoding and the continuous encoding.

4. The optical diffractive processing unit of claim 2 , wherein the diffractive connection layer is implemented by optical diffraction.

5. The optical diffractive processing unit of claim 2 , wherein the optical summation layer is implemented by optical coherence.

6. The optical diffractive processing unit of claim 2 , wherein the optical non-linearity layer is implemented by a programmable detection module.

7. The optical diffractive processing unit of claim 2 , wherein the single-layer diffractive layers are connected sequentially.

8. The optical diffractive processing unit of claim 1 , wherein the optical neural network comprises a three-layer optoelectronic diffractive deep neural network (D 2 NN).

9. The optical diffractive processing unit of claim 1 , wherein the optical neural network comprises recurrent modules;

an output of each recurrent module is a recurrent state;

an input of each recurrent module comprises an output state of a previous recurrent module and sequential input data;

each recurrent module comprises multiple states inside; and

another diffractive neural network architecture is connectable to a recurrent neural network during an inference process.