A training system for training a machine learning model such as a neural network may have a different configuration and/or hardware components than a target device that employs the trained neural network. For example, the training system may use a higher precision format to represent neural network parameters than the target device. In another example, the target device may use analog and digital processing hardware to compute an output of the neural network whereas the training system may have used only digital processing hardware to train the neural network. The difference in configuration and/or hardware components of the target device may introduce quantization error into parameters of the neural network, and thus affect performance of the neural network on the target device. Described herein is a training system that trains a neural network for use on a target device that reduces loss in performance resulting from quantization error.

Embodiments of the present disclosure are directed toward techniques and apparatus comprising at least one layer of an ONN that includes an optical matrix multiplier provided in a semiconductor substrate to receive a plurality of optical signal inputs and to linearly transform the plurality of optical signal inputs into a plurality of optical signal outputs. The optical matrix multiplier comprises one or more 2×2 unitary optical matrices optically interconnected to implement a singular value decomposition (SVD) of a matrix, and a nonlinear optical device coupled with the optical matrix multiplier in the semiconductor substrate, to receive the optical signal outputs and to provide an optical output that is generated in a nonlinear manner in response to the optical signal outputs of the optical matrix multiplier reaching saturation or attenuation. Additional embodiments may be described and claimed.

The method includes receiving, at a first optical communication device, feedback information on training of a neural network from at least one second optical communication device, the neural network configured to process a signal received from the first optical communication device, the feedback information at least including a training performance indication for training of the neural network conducted at the at least one second optical communication device; updating, based on the feedback information, a first initial parameter value set for the neural network maintained at the first optical communication device, to obtain a second initial parameter value set for the neural network; and transmitting the second initial parameter value set to at least one further second optical communication device, for training of the neural network to be conducted at the at least one further second optical communication device based on the second initial parameter value set.

Systems and methods for performing signed matrix operations using a linear photonic processor are provided. The linear photonic processor is formed as an array of first amplitude modulators and second amplitude modulators, the first amplitude modulators configured to encode elements of a vector into first optical signals and the second amplitude modulators configured to encode a product between the vector elements and matrix elements into second optical signals. An apparatus may be used to implement a signed value of an output of the linear processor. The linear photonic processor may be configured to perform matrix-vector and/or matrix-matrix operations.

A photonic processor uses light signals and a residue number system (RNS) to perform calculations. The processor sums two or more values by shifting the phase of a light signal with phase shifters and reading out the summed phase with a coherent detector. Because phase winds back every 2π radians, the photonic processor performs addition modulo 2π. A photonic processor may use the summation of phases to perform dot products and correct erroneous residues. A photonic processor may use the RNS in combination with a positional number system (PNS) to extend the numerical range of the photonic processor, which may be used to accelerate homomorphic encryption (HE)-based deep learning.

A hybrid neuromorphic computing device is provided, in which artificial neurons include light-emitting devices that provide weighted sums of inputs as light output. The output is detected by a photodetector and converted to an electrical output. Each neuron may receive output from one or more other neurons as initial input. Interconnects between neurons may be optical, electrical, or a combination thereof. The neurons also may provide imaging sensor and/or display capabilities.

Methods, systems, and apparatus for performing convolutional computations using an optical system. In some aspects computations for a neural network are performed in a digital domain using electronic circuitry, the neural network including a convolutional layer. Input data for the convolutional layer of the neural network is obtained, and a convolution or correlation computation on the input data in an analog domain using an optical correlator module is performed to generate an optical correlator module output. Based on the optical correlator module output, data is processed through additional layers of the neural network in the digital domain using the electronic circuitry.

To realize a reservoir computing system easily implemented as hardware, provided is a reservoir computing system including a reservoir operable to output an inherent output signal in response to an input signal. An input node is operable to supply the reservoir with an input signal corresponding to input data, and an output node is operable to output an output value corresponding to an output signal that is output by the reservoir in response to the input data. An adaptive filter is operable to output output data based on a result obtained by weighting a plurality of the output values output from the output node at a plurality of timings with a plurality of weights. Also provided are a learning method and a computer program product.

Most artificial neural networks are implemented electronically using graphical processing units to compute products of input signals and predetermined weights. The number of weights scales as the square of the number of neurons in the neural network, causing the power and bandwidth associated with retrieving and distributing the weights in an electronic architecture to scale poorly. Switching from an electronic architecture to an optical architecture for storing and distributing weights alleviates the communications bottleneck and reduces the power per transaction for much better scaling. The weights can be distributed at terabits per second at a power cost of picojoules per bit (versus gigabits per second and femtojoules per bit for electronic architectures). The bandwidth and power advantages are even better when distributing the same weights to many optical neural networks running simultaneously.

An optical computing device and a computing method are provided, to provide an optical Ising machine with high operation efficiency. The optical computing device includes a first spin array, an optical feedback network, and a second spin array, where the optical feedback network is separately connected to the first spin array and the second spin array. The first spin array may receive a first group of signals including N optical pulses or N electrical signals, and generate a first group of spin signals including N spin signals. The optical feedback network may receive the first group of spin signals, and generate, based on the first group of spin signals and specified first data, a first group of feedback signals including N feedback signals. The first spin array and the second spin array may process a plurality of signals in parallel, to improve computation efficiency of the optical computing device.