TY - GEN
T1 - EVA2
T2 - 45th ACM/IEEE Annual International Symposium on Computer Architecture, ISCA 2018
AU - Buckler, Mark
AU - Bedoukian, Philip
AU - Jayasuriya, Suren
AU - Sampson, Adrian
N1 - Publisher Copyright:
© 2018 IEEE.
PY - 2018/7/19
Y1 - 2018/7/19
N2 - Hardware support for deep convolutional neural networks (CNNs) is critical to advanced computer vision in mobile and embedded devices. Current designs, however, accelerate generic CNNs; they do not exploit the unique characteristics of real-time vision. We propose to use the temporal redundancy in natural video to avoid unnecessary computation on most frames. A new algorithm, activation motion compensation, detects changes in the visual input and incrementally updates a previously-computed activation. The technique takes inspiration from video compression and applies well-known motion estimation techniques to adapt to visual changes. We use an adaptive key frame rate to control the trade-off between efficiency and vision quality as the input changes. We implement the technique in hardware as an extension to state-of-the-art CNN accelerator designs. The new unit reduces the average energy per frame by 54%, 62%, and 87% for three CNNs with less than 1% loss in vision accuracy.
AB - Hardware support for deep convolutional neural networks (CNNs) is critical to advanced computer vision in mobile and embedded devices. Current designs, however, accelerate generic CNNs; they do not exploit the unique characteristics of real-time vision. We propose to use the temporal redundancy in natural video to avoid unnecessary computation on most frames. A new algorithm, activation motion compensation, detects changes in the visual input and incrementally updates a previously-computed activation. The technique takes inspiration from video compression and applies well-known motion estimation techniques to adapt to visual changes. We use an adaptive key frame rate to control the trade-off between efficiency and vision quality as the input changes. We implement the technique in hardware as an extension to state-of-the-art CNN accelerator designs. The new unit reduces the average energy per frame by 54%, 62%, and 87% for three CNNs with less than 1% loss in vision accuracy.
KW - Application specific integrated circuits
KW - Computer architecture
KW - Computer vision
KW - Convolutional neural networks
KW - Hardware acceleration
KW - Video compression
UR - http://www.scopus.com/inward/record.url?scp=85055871992&partnerID=8YFLogxK
UR - http://www.scopus.com/inward/citedby.url?scp=85055871992&partnerID=8YFLogxK
U2 - 10.1109/ISCA.2018.00051
DO - 10.1109/ISCA.2018.00051
M3 - Conference contribution
AN - SCOPUS:85055871992
T3 - Proceedings - International Symposium on Computer Architecture
SP - 533
EP - 546
BT - Proceedings - 2018 ACM/IEEE 45th Annual International Symposium on Computer Architecture, ISCA 2018
PB - Institute of Electrical and Electronics Engineers Inc.
Y2 - 2 June 2018 through 6 June 2018
ER -