A multiplication reduction technique with near-zero approximation for embedded learning in IoT devicesVise andre og tillknytning
2017 (engelsk)Inngår i: International System on Chip Conference, IEEE Computer Society , 2017, s. 102-107Konferansepaper (Fagfellevurdert)
Abstract [en]
This paper presents a multiplication reduction technique through near-zero approximation, enabling embedded learning in resource-constrained IoT devices. The intrinsic resilience of neural network and the sparsity of data are identified and utilized. Based on the analysis of leading zero counting and adjustable threshold, intentional approximation is applied to reduce near-zero multiplications. By setting the threshold of the multiplication result to 2-5 and employing ReLU as the neuron activation function, the sparsity of the CNN model can reach 75% with negligible loss in accuracy when recognizing the MNIST data set. Corresponding hardware implementation has been designed and simulated in UMC 65nm process. It can achieve more than 70% improvement of energy efficiency with only 0.37% area overhead of a 256 Multiply-Accumulator array.
sted, utgiver, år, opplag, sider
IEEE Computer Society , 2017. s. 102-107
Emneord [en]
Energy efficiency, Hardware, Programmable logic controllers, Area overhead, CNN models, Data set, Embedded learning, Hardware implementations, Multiply accumulators, Neuron activation function, Reduction techniques, Internet of things
HSV kategori
Identifikatorer
URN: urn:nbn:se:kth:diva-216527DOI: 10.1109/SOCC.2016.7905445ISI: 000403576000018Scopus ID: 2-s2.0-85019109259ISBN: 9781509013661 (tryckt)OAI: oai:DiVA.org:kth-216527DiVA, id: diva2:1161802
Konferanse
29th IEEE International System on Chip Conference, SOCC 2016, 6 September 2016 through 9 September 2016
Merknad
QC 20171201
2017-12-012017-12-012017-12-01bibliografisk kontrollert