A mixed signal architecture for convolutional neural networks

Lou, Qiuwen; Pan, Chenyun; McGuiness, John C.; Horváth, András; Naeemi, Azad; Niemier, Michael; Hu, Xiaobo Sharon

doi:10.48550/arxiv.1811.02636

Cited by 2 publications

(11 citation statements)

References 45 publications

Supporting

Mentioning

Contrasting

Order By: Relevance

“…Lou et al proposed using standard, purely chargebased CeNN cells with weighted, programmable Operational Transconductance Amplifier-based current sources (OTAs) to produce an in-hardware implementation of a convolutional neural network [1], [5]. Using CeNN weight template schemes, the Rectified Linear Unit (ReLU) activation function and pooling function were approximated so that all but the fullyconnected output layer of a CoNN could be implemented via simple CeNN cells.…”

Section: Convolutional Network With Irmen Cellsmentioning

confidence: 99%

“…Each stage corresponds to one CeNN operation, represented by a specific CeNN template imposed upon the data as it is slowly transformed from the initial input to the final output. Each stage of computation is implemented by one of a pair of identical IRMEN CeNNs similar to the structure in [5]. One of the pair provides the input, obtained from the previous stage and locally stored, to the other of the pair for processing and subsequent storing (see Fig.…”

Section: A Coupled Cenns For Memory and Computingmentioning

confidence: 99%

“…The structure of the CoNN to be implemented via IRMEN CeNNs is shown in Fig. 5, is identical to one of the test cases in [5]. In the interest of time the training of the weights was implemented using TensorFlow with a custom transfer function representing a close numerical approximation to the hysteresis curve of the IRMENs' nanomagets.…”

Section: A Simulation Setupmentioning

confidence: 99%

“…The dot-product has been successfully implemented in hardware in various ways [1]- [4]. Using these dot-product circuits, a cellular neural network-based (CeNN) convolutional neural network (CoNN) accelerator was designed in [5]. We propose a CeNN cell based on recently proposed spintronic elements with a high energy-efficiency that can be used as analog memory with a built-in activation function.…”

Section: Introductionmentioning

confidence: 99%

“…We propose a CeNN cell based on recently proposed spintronic elements with a high energy-efficiency that can be used as analog memory with a built-in activation function. The performance of these cells is simulated in a CeNN-accelerated CoNN performing image classification based on [5]. The spintronic cells significantly reduce the energy and time consumption relative to their charge-based counterparts, needing only ≈ 100 pJ and ≈ 42 ns to compute all but the final fully-connected CoNN layer while maintaining a high accuracy.…”

Section: Introductionmentioning

confidence: 99%

See 4 more Smart Citations

Nonvolatile Spintronic Memory Cells for Neural Networks

Stephan¹,

Lou²,

Niemier³

et al. 2019

Preprint

Self Cite

View full text Add to dashboard Cite

A new spintronic nonvolatile memory cell analogous to 1T DRAM with non-destructive read is proposed. The cells can be used as neural computing units. A dual-circuit neural network architecture is proposed to leverage these devices against the complex operations involved in convolutional networks. Simulations based on HSPICE and Matlab were performed to study the performance of this architecture when classifying images as well as the effect of varying the size and stability of the nanomagnets. The spintronic cells outperform a purely chargebased implementation of the same network, consuming ≈ 100 pJ total per image processed.

show abstract

Section: Convolutional Network With Irmen Cellsmentioning

confidence: 99%

Section: A Coupled Cenns For Memory and Computingmentioning

confidence: 99%

Section: A Simulation Setupmentioning

confidence: 99%

Section: Introductionmentioning

confidence: 99%

Section: Introductionmentioning

confidence: 99%

See 3 more Smart Citations

Nonvolatile Spintronic Memory Cells for Neural Networks

Stephan¹,

Lou²,

Niemier³

et al. 2019

Preprint

Self Cite

View full text Add to dashboard Cite

show abstract

Convolutional Neural Networks Utilizing Multifunctional Spin-Hall MTJ Neurons

Stephan,

Koester

2019

Preprint

View full text Add to dashboard Cite

We propose a new network architecture for standard spin-Hall magnetic tunnel junction-based spintronic neurons that allows them to compute multiple critical convolutional neural network functionalities simultaneously and in parallel, saving space and time. An approximation to the Rectified Linear Unit transfer function and the local pooling function are computed simultaneously with the convolution operation itself. A proof-of-concept simulation is performed on the MNIST dataset, achieving up to 98% accuracy at a cost of less than 1 nJ for all convolution, activation and pooling operations combined. The simulations are remarkably robust to thermal noise, performing well even with very small magnetic layers.

show abstract

A mixed signal architecture for convolutional neural networks

Cited by 2 publications

References 45 publications

Nonvolatile Spintronic Memory Cells for Neural Networks

Nonvolatile Spintronic Memory Cells for Neural Networks

Convolutional Neural Networks Utilizing Multifunctional Spin-Hall MTJ Neurons

Contact Info

Product

Resources

About