Acoustooptical matrix-vector product processor: implementation issues

Mechanical Neural Networks with Explicit and Robust Neurons

Mechanical computing provides an information processing method to realize sensing-analyzing-actuation integrated mechanical intelligence and, when combined with neural networks, can be more efficient for data-rich cognitive tasks. The requirement of solving implicit and usually nonlinear equilibrium equations of motion in training mechanical neural networks makes computation challenging and costly. Here, an explicit mechanical neuron is developed of which the response can be directly determined without the need of solving equilibrium equations. A training method is proposed to ensure the robustness of the neuron, i.e., insensitivity to defects and perturbations. The explicitness and robustness of the neurons facilitate the assembly of various network structures. Two exemplified networks, a robust mechanical convolutional neural network and a mechanical recurrent neural network with long short-term memory capabilities for associative learning, are experimentally demonstrated. The introduction of the explicit and robust mechanical neuron streamlines the design of mechanical neural networks fulfilling robotic matter with a level of intelligence.

Key Technologies of Photonic Artificial Intelligence Chip Structure and Algorithm

Artificial intelligence chips (AICs) are the intersection of integrated circuits and artificial intelligence (AI), involving structure design, algorithm analysis, chip fabrication and application scenarios. Due to their excellent ability in data processing, AICs show a long-term industrial prospect in big data services, cloud centers, etc. However, with the conceivable exhaustion of Moore’s Law, the size of traditional electronic AICs (EAICs) is gradually approaching the limit, and an architectural update is highly required. Photonic artificial intelligence chips (PAIC) utilize light beam propagation in the silicon waveguide, contributing to a high parallelism configuration, fast calculation speed and low latency. Due to light manipulation, PAICs perform well in anti-electromagnetic interference and energy conservation. This invited paper summarized the recent research on PAICs. The characteristics of different hardware structures are discussed. The current widely used training algorithm is given and the Photonic Design Automatic (PDA) simulation platform is introduced. In addition, the authors’ related work on PAICs is presented and we believe that PAICs may play a critical role in the deployment of data processing technology.

Fully reconfigurable coherent optical vector-matrix multiplication

Optics is a promising platform in which to help realize the next generation of fast, parallel, and energy-efficient computation. We demonstrate a reconfigurable free-space optical multiplier that is capable of over 3000 computations in parallel, using spatial light modulators with a pixel resolution of only 340×340. This enables vector-matrix multiplication and parallel vector-vector multiplication with vector size of up to 56. Our design is, to the best of our knowledge, the first to simultaneously support optical implementation of reconfigurable, large-sized, and real-valued linear algebraic operations. Such an optical multiplier can serve as a building block of special-purpose optical processors such as optical neural networks and optical Ising machines.

Carry-free vector-matrix multiplication on a dynamically reconfigurable optical platform

Applying the parallelism of optical computing, we present a novel method of vector-matrix multiplication (VMM) based on a new optical computing platform, the ternary optical computer, which can reconfigure any two-input trivalued logic optical processor at runtime, according to the decrease-radix design principle. In this work, we investigate a novel optical VMM (OVMM) using five logic operations with the modified signed-digit (MSD) number system. To simplify the computation process, we realize a carry-free optical addition in three steps, which is independent of the length of the operands. And a new implementation method is proposed that can be used to realize the MSD multiplication in parallel. Based on the generation of partial products in parallel and the binary-addition-tree algorithm, the multiplication can be implemented with the MSD addition. Our initial experiments have been performed to verify the proposed OVMM method. The results show that the proposed method of OVMM is feasible and correct.

Two-dimensional multi-channel acousto-optic diffraction

In this paper, we have investigated the interaction between acoustic waves generated by an incident light beam propagating through two-dimensional transducers. A set of coupled mode equations for two-dimensional multi-channel acousto-optic inter-modulation are derived, and the solutions are obtained. A new type of two-dimensional multi-channel Bragg acousto-optic modulator is designed by analyzing two-dimensional multi-channel acousto-optic inter-modulation based on the coupled mode theory, and its performance has also been experimentally characterized. The experimental results are in accordance with our theoretical analysis.

Modified engagement method for matrix operation

We describe a modified engagement method for matrix operation based on a two-dimensional crossed-ring interconnection network. Our method incorporates fewer steps than that reported by Bocker et al. [Appl. Opt. 22, 804 (1983)], and its performance is found to be the most efficient (minimum steps) in comparison with other systolic and/or engagement methods for matrix operation. Thus, it may be helpful for other optical and electronic implementations of matrix operations. One compact optoelectronic integrity approach for implementing the modified engagement method is briefly described.

Modified direct twos-complement parallel array multiplication algorithm for complex matrix operation

A direct twos-complement parallel array multiplication algorithm is introduced and modified for digital optical numerical computation. The modified version overcomes the problems encountered in the conventional optical twos-complement algorithm. In the array, all the summands are generated in parallel, and the relevant summands having the same weights are added simultaneously without carries, resulting in the product expressed in a mixed twos-complement system. In a two-stage array, complex multiplication is possible with using four real subarrays. Furthermore, with a three-stage array architecture, complex matrix operation is straightforwardly accomplished. In the experiment, parallel two-stage array complex multiplication with liquid-crystal panels is demonstrated.

Negative binary arithmetic algorithms for digital parallel optical computation

Based on a negative binary number system, an algorithm with weighted-shifted addition, parallel-array multiplication, and a two-stage-array complex operation is proposed to carry out the multiplication of two complex numbers. The complex multiplication is performed without signs, carries, and recoding. The algorithm is suitable for optical implementation, and an optical parallel architecture is suggested. The experimental result is also given.

Optical expanders with applications in optical computing

An optical system called the optical expander is described and investigated. The optical expander electro-optically expands an optical Boolean pattern encoded in d bits into an optical pattern of size N bits. It is assumed that d is equal to c log(2)N for some constant c, and each expanded pattern is orthogonal to the others. Two different architectures to implement the optical expander are described: one uses an optical matrix-vector multiplier and an array of N threshold devices; the other uses log(2)N novel reflection-transmission switching cells. These architectures are analyzed in terms of size, energy requirement, and speed. The optical expander described utilizes high-speed and high-space-bandwidth-product connections that are provided by optical beams in three dimensions. Potential applications, holographic memory, and message routing systems are also discussed.

Matrix-vector multiplication using digital partitioning for more accurate optical computing

Digital partitioning offers a flexible means of increasing the accuracy of an optical matrix-vector processor. This algorithm can be implemented with the same architecture required for a purely analog processor, which gives optical matrix-vector processors the ability to perform high-accuracy calculations at speeds comparable with or greater than electronic computers as well as the ability to perform analog operations at a much greater speed. Digital partitioning is compared with digital multiplication by analog convolution, residue number systems, and redundant number representation in terms of the size and the speed required for an equivalent throughput as well as in terms of the hardware requirements. Digital partitioning and digital multiplication by analog convolution are found to be the most efficient algorithms if coding time and hardware are considered, and the architecture for digital partitioning permits the use of analog computations to provide the greatest throughput for a single processor. To our knowledge this is the first study to propose the use of digital partitioning for optical matrix processing.