2019
DOI: 10.1063/1.5108912
|View full text |Cite
|
Sign up to set email alerts
|

Novel frontier of photonics for data processing—Photonic accelerator

Abstract: In the emerging Internet of things cyber-physical system-embedded society, big data analytics needs huge computing capability with better energy efficiency. Coming to the end of Moore’s law of the electronic integrated circuit and facing the throughput limitation in parallel processing governed by Amdahl’s law, there is a strong motivation behind exploring a novel frontier of data processing in post-Moore era. Optical fiber transmissions have been making a remarkable advance over the last three decades. A reco… Show more

Help me understand this report

Search citation statements

Order By: Relevance

Paper Sections

Select...
5

Citation Types

0
89
0

Year Published

2020
2020
2024
2024

Publication Types

Select...
6
2
1

Relationship

1
8

Authors

Journals

citations
Cited by 169 publications
(89 citation statements)
references
References 99 publications
0
89
0
Order By: Relevance
“…The cascaded layers of phase-only transmissive arrays are trained by deep diffractive neuron networks [8]. Compared to the waveguide based integrated photonic processors [31][32][33], the metasystem architecture offers three orders of magnitude higher throughput of vector-bymatrix multiplication (Supplementary Note S1) [34]. The metamaterial manifested weight element density, combined with diffraction strengthen inter-layer connectivity, enables the passive system to accomplish machine learning tasks of spatial pattern classification (Fig.…”
Section: Introductionmentioning
confidence: 99%
“…The cascaded layers of phase-only transmissive arrays are trained by deep diffractive neuron networks [8]. Compared to the waveguide based integrated photonic processors [31][32][33], the metasystem architecture offers three orders of magnitude higher throughput of vector-bymatrix multiplication (Supplementary Note S1) [34]. The metamaterial manifested weight element density, combined with diffraction strengthen inter-layer connectivity, enables the passive system to accomplish machine learning tasks of spatial pattern classification (Fig.…”
Section: Introductionmentioning
confidence: 99%
“…Different from traditional recurrent neural network algorithm, RC's connection weights of the input layer and the reservoir are fixed during training procedure, and only the readout weights need to be trained through a straightforward linear regression method. Since the training procedure only involves the readout layer without changing the reservoir connection, RC possesses such unique virtues as fast training process and easy hardware implementation [4]. Traditionally, RC is implemented based on a spatially distributed network [5]- [7], where the reservoir is constituted by numerous interconnected nonlinear nodes and the reservoir states (the states of nonlinear nodes) are obtained in parallel.…”
Section: Introductionmentioning
confidence: 99%
“…In recent years, though relevant researches on SL-based time-delay RC have made significant progresses [19]- [31], the data processing rate still needs to be further improved to meet the requirement of fast information processing [4]. In a SL-based time-delay RC system, an original data is firstly sampled and held for a period T. Next, the resulted signal is multiplied by a temporal mask to increase the dimensionality of the state space.…”
Section: Introductionmentioning
confidence: 99%
“…Recently, photonic approaches to decision-making problems have been intensively studied using single photons 12 , 13 , chaotic lasers 14 , 15 , and entangled photons 16 . Related topics are discussed in recent review articles 1 , 17 , 18 .…”
Section: Introductionmentioning
confidence: 99%