A Survey on Silicon Photonics for Deep Learning

Sunny, Febin; Taheri, Ebadollah; Nikdast, Mahdi; Pasricha, Sudeep

doi:10.48550/arxiv.2101.01751

Cited by 2 publications

(5 citation statements)

References 0 publications

Supporting

Mentioning

Contrasting

Order By: Relevance

“…Each connection between the neurons is assigned a weight that represents its synaptic plasticity and each neuron is tasked with a multiply-and-accumulate (MAC) operation followed by passing the resultant output through a non-linear activation function (f N AU ). By introducing non-linearity in the network, the activation functions (e.g., sigmoid, tanh, and Rectified Linear Unit) enable the DNNs to learn complex non-linear relationships [6]. During each training iteration, the weight of each connection in a DNN is incrementally updated to minimize the loss function that quantifies the difference between the expected and the obtained DNN output.…”

Section: B Mzi-based Coherent Ipnnsmentioning

confidence: 99%

“…Integrated photonic neural networks (IPNNs) based on silicon photonics can expedite extensive linear operations (i.e., matrix multiplication) in DNNs [6]. By taking advantage of the natural parallelism in photonics, computations in IPNNs can be performed in parallel, hence reducing the complexity of matrix-vector multiplication in DNNs from O(N 2 ) to approximately O(1) [7].…”

Section: Introductionmentioning

confidence: 99%

“…1 presents a hierarchical view of a multi-layer perceptron (MLP)-based IPNN. On the other hand, noncoherent neurons employ the Broadcast-and-Weight configuration to manipulate the power of multiplewavelength optical signals [6]. While noncoherent neurons support wavelength-division multiplexing, they suffer from inter-channel crosstalk and the dependency between the input and output wavelengths necessitates energy-intensive wavelength conversion steps [6].…”

Section: Introductionmentioning

confidence: 99%

“…On the other hand, noncoherent neurons employ the Broadcast-and-Weight configuration to manipulate the power of multiplewavelength optical signals [6]. While noncoherent neurons support wavelength-division multiplexing, they suffer from inter-channel crosstalk and the dependency between the input and output wavelengths necessitates energy-intensive wavelength conversion steps [6]. As a result, recent commercial efforts on developing photonic neural networks have focused on using coherent neurons to expedite matrix multiplication at a low energy cost [15].…”

Section: Introductionmentioning

confidence: 99%

See 3 more Smart Citations

Characterizing Coherent Integrated Photonic Neural Networks under Imperfections

Banerjee¹,

Nikdast²,

Chakrabarty³

2022

Preprint

Self Cite

View full text Add to dashboard Cite

Integrated photonic neural networks (IPNNs) are emerging as promising successors to conventional electronic AI accelerators as they offer substantial improvements in computing speed and energy efficiency. In particular, coherent IPNNs use arrays of Mach-Zehnder interferometers (MZIs) for unitary transformations to perform energy-efficient matrix-vector multiplication. However, the underlying MZI devices in IPNNs are susceptible to uncertainties stemming from optical lithographic variations and thermal crosstalk and can experience imprecisions due to non-uniform MZI insertion loss and quantization errors due to low-precision encoding in the tuned phase angles. In this paper, we, for the first time, systematically characterize the impact of such uncertainties and imprecisions (together referred to as imperfections) in IPNNs using a bottom-up approach. We show that their impact on IPNN accuracy can vary widely based on the tuned parameters (e.g., phase angles) of the affected components, their physical location, and the nature and distribution of the imperfections. To improve reliability measures, we identify critical IPNN building blocks that, under imperfections, can lead to catastrophic degradation in the classification accuracy. We show that under multiple simultaneous imperfections, the IPNN inferencing accuracy can degrade by up to 46%, even when the imperfection parameters are restricted within a small range. Our results also indicate that the inferencing accuracy is sensitive to imperfections affecting the MZIs in the linear layers next to the input layer of the IPNN.

show abstract

Section: B Mzi-based Coherent Ipnnsmentioning

confidence: 99%

Section: Introductionmentioning

confidence: 99%

Section: Introductionmentioning

confidence: 99%

Section: Introductionmentioning

confidence: 99%

See 2 more Smart Citations

Characterizing Coherent Integrated Photonic Neural Networks under Imperfections

Banerjee¹,

Nikdast²,

Chakrabarty³

2022

Preprint

Self Cite

View full text Add to dashboard Cite

show abstract

“…The researchers are attempting to find out other methods to solve electronic defects. One of the most promising answers to solve problems with data transportation is photonic interconnects or all-optical computing systems [3]. When GPUs are used for applications such as facial recognition, image classification, deep learning applications, GPUs will transform the optical information into electrical signals.…”

Section: Introductionmentioning

confidence: 99%

Image compression in all-optical domain using one 6×6 multimode interference coupler

Thuy¹,

Thanh

2023

Vietnam J. Sci. Technol.

View full text Add to dashboard Cite

We present a new method for image compression using the Haar wavelet transform (HWT) in all-optical domain. The proposed architecture is based on the optical Haar wavelet transform using only one 6×6 multimode interference (MMI) coupler. By locating the positions of input and output waveguides and optimizing design length of the MMI, the expected matrix of the Haar transform is obtained. The new hardware architecture is suitable for directly integrating with digital cameras for image processing. The processing of images therefore is at very high speed. Our method can also be applied to data compression in big data analytics. Our structure can provide a large fabrication tolerance which is compatible with the CMOS existing technology. Our simulations show that the length variation of ± 2 µm still keeps the output powers unchanged. We have simulate successfully the use of the proposed HWT to compress cameraman image with the compressed ratios of 20, 30, 50 % with MSE and PSNR from 0.1 - 0.3 and 62 – 67 dB, respectively.

show abstract

A Survey on Silicon Photonics for Deep Learning

Cited by 2 publications

References 0 publications

Characterizing Coherent Integrated Photonic Neural Networks under Imperfections

Characterizing Coherent Integrated Photonic Neural Networks under Imperfections

Image compression in all-optical domain using one 6×6 multimode interference coupler

Contact Info

Product

Resources

About