An Efficient Multilayer Spiking Convolutional Neural Network Processor for Object Recognition With Low Bitwidth and Channel-Level Parallelism

Previous studies have shown that the event-driven multilayer spiking convolutional neural network (SCNN) can reduce computational complexity largely while keeping accurate. To fully utilize the advantages of SCNN, this brief proposed an efficient multilayer SCNN processor for object recognition. The...

Celý popis

Uloženo v:
Podrobná bibliografie
Vydáno v:IEEE transactions on circuits and systems. II, Express briefs Ročník 69; číslo 12; s. 5129 - 5133
Hlavní autoři: Feng, Lichen, Zhang, Yueqi, Zhu, Zhangming
Médium: Journal Article
Jazyk:angličtina
Vydáno: New York IEEE 01.12.2022
The Institute of Electrical and Electronics Engineers, Inc. (IEEE)
Témata:
Artificial neural networks
Channel-level parallel
Convolutional neural networks
Encoding
Image classification
Microprocessors
Multilayers
Neurons
Nonhomogeneous media
Object recognition
Parallel operation
Random access memory
rank-order coding
sparse event
spike-centric
Spiking
spiking convolutional neural network
Support vector machines
ISSN:1549-7747, 1558-3791
On-line přístup:Získat plný text
Tagy: Přidat tag
Žádné tagy, Buďte první, kdo vytvoří štítek k tomuto záznamu!
Popis
Shrnutí:Previous studies have shown that the event-driven multilayer spiking convolutional neural network (SCNN) can reduce computational complexity largely while keeping accurate. To fully utilize the advantages of SCNN, this brief proposed an efficient multilayer SCNN processor for object recognition. The interconnection between spiking layers is implemented for the first time. The rank-order coding with mutual and lateral inhibitions enables sparse event transmission. By further combining the spike-centric membrane potential update, channel-level parallel operation, and the low bitwidths of synapse weights and potentials, the proposed design achieves 500 classifications/s, and 68 uJ/classification for recognizing images with <inline-formula> <tex-math notation="LaTeX">160\mathbf {\times }250 </tex-math></inline-formula> resolution, which is superior to the recent works.
Bibliografie:ObjectType-Article-1
SourceType-Scholarly Journals-1
ObjectType-Feature-2
content type line 14
ISSN:1549-7747
1558-3791
DOI:10.1109/TCSII.2022.3207989