Abstract:Photonic neural networks (PNNs), as emerging analog computing paradigms, have the potential to break both the power consumption wall and processing speed boundary of digital computing architectures. However, most reported PNNs were mainly focused on real-valued regions, where phases of optical fields were only utilized as indirect information to affect optical amplitudes. Moreover, the quantization of phase information at each node of multi-layer PNNs is disregarded. In this work, we propose and demonstrate a Clements-enhanced complex-valued mesh (CCM) with balanced detection units (BDUs), which can capture both the amplitude and phase information at one layer and pass them on to the next layer. Besides, by introducing $N/2$ additional phase shifters to the Clements architecture, the proposed CCM possesses a complete characterization ability for arbitrary N-order complex unitary matrices. In this way, both the input vectors and the weighted matrices can be mapped into complex-valued regions and further implemented on the proposed CCM, in which the corresponding computing density is increased by four times with the comparison of real-valued meshes. To validate the proposed CCM with BDUs, we also fabricated a silicon-based on-chip 4×4 mesh with a footprint of 7.86 mm $^{2}$ and experimentally verified its performances on a two-class classification task of King-Rook vs. King (KRK) in chess and a ten-class classification of handwritten digits from MNIST database, yielding tested accuracies of 94.06 $\%$ and 91.75 $\%$ , respectively. The proposed passive on-chip CCM together with BDUs might provide another solution for next-generation artificial intelligence hardware with enhanced performances.

A CMOS Implementation for Radial Basis Function Networks

Analog design of a radial basis function network

Hardware implementation of a combined radial basis function neural network

Hardware implementation of radial-basis neural networks with Gaussian activation functions on FPGA

Implementation of SoC of Soft Measurement Algorithm Based on RBF Neural Network

Analog/digital circuit simplification for Hopfield neural network

A Proposal for Energy-Efficient Cellular Neural Network based on Spintronic Devices

Clements-Enhanced Complex-Valued Coherent Mesh with Balanced Detection Units for Photonic Neural Networks

An SRAM-Based Reconfigurable Cognitive Computation Matrix for Sensor Edge Applications

A Low-Power Analog Integrated Euclidean Distance Radial Basis Function Classifier

An Improved RRAM-Based Binarized Neural Network with High Variation-Tolerated Forward/Backward Propagation Module

Dot-product computation and logistic regression with 2D hexagonal-Boron Nitride (h-BN) memristor arrays

Design of a new CMOS Low-Power Analogue Neuron

Radial basis function neural networks: a topical state-of-the-art survey

Universal Approximation Using Radial-Basis-Function Networks

Multi-Kernel Fusion for RBF Neural Networks

A CMOS implementation of FitzHugh-Nagumo neuron model

A New Hardware Implementation Approach Of Bnns Based On Nonlinear 2t2r Synaptic Cell

A Novel Multiplier For Achieving The Programmability Of Cellular Neural Network

Radial Basis Operator Networks

Improved blocks for CMOS analog neuro-fuzzy network