DATA PRUNING METHOD FOR LIGHTWEIGHT DEEP-LEARNING HARDWARE DEVICE

WIPO Patent Application WO/2024/135860

A1

A data pruning method for a lightweight deep-learning hardware device is provided. A deep-learning computation method according to an embodiment of the present invention comprises: pruning a data stream output from a previous layer of a deep learning model, on the basis of the unit of channels of a channel unit; binding the pruned data stream by multiple channel units; and transferring the bound data stream to a next layer of the deep learning model. Accordingly, the method enables customized pruning that takes NPU specifications into account to perform filter pruning capable of omitting and reducing computation in lightweight deep-learning hardware applied to edge devices, resulting in increased hardware space utilization and minimization of the unnecessary computation amount.

WO/2023/022505	METHOD FOR PERFORMING BNN OPERATION THROUGH IN-MEMORY COMPUTING USING DRAM, AND DEVICE USING SAME
WO/2023/056779	COMPUTING-IN-MEMORY EDRAM ACCELERATOR FOR CONVOLUTIONAL NEURAL NETWORK
WO/2023/027368	EXECUTION ENGINE OPTIMIZATION METHOD, EXECUTION ENGINE OPTIMIZATION DEVICE, AND EXECUTION ENGINE OPTIMIZATION SYSTEM

LEE, Sang Seol (12-240 Sinhyeon-ro,Opo-eup, Gwangju-si, Gyeonggi-do, KR)
LEE, Eun Chong (16-16 Saemal-ro 15-gil,Songpa-gu, Seoul, KR)
KIM, Kyung Ho (25 Irwon-ro 14-gil,Gangnam-gu, Seoul, KR)

PCT/KR2022/020663

June 27, 2024

December 19, 2022

Click for automatic bibliography generation  

KOREA ELECTRONICS TECHNOLOGY INSTITUTE (Bundang-gu, Seongnam-si, Gyeonggi-do, KR)

G06N3/063; G06N3/08

NAM, Choong Woo (Gangnam-gu, Seoul, KR)

View/Download PDF      PDF Help

Previous Patent: ANTENNA, ANTENNA DEVICE, AND ANTENNA UNIT

Next Patent: DEEP LEARNING TRAINING METHOD APPLYING VARIABLE DATA REPRESENTATION TYPE AND MOBILE DEVICE APPLYING ...