US 12,254,407 B2
Storage and inference method for deep-learning neural network
Yen-Chin Liao, Hsinchu (TW)
Assigned to HUITONG INTELLIGENCE COMPANY LIMITED, Taipei (TW)
Filed by HUITONG INTELLIGENCE COMPANY LIMITED, Taipei (TW)
Filed on Aug. 16, 2021, as Appl. No. 17/402,970.
Claims priority of application No. 110119690 (TW), filed on May 31, 2021.
Prior Publication US 2022/0405577 A1, Dec. 22, 2022
Int. Cl. G06N 3/045 (2023.01); G06N 3/08 (2023.01)
CPC G06N 3/08 (2013.01) [G06N 3/045 (2023.01)] 8 Claims
OG exemplary drawing
 
1. A storage and inference method for a deep-learning neural network, comprising steps:
establishing a plurality of dummy nodes in a first artificial neural network to form a second artificial neural network;
storing model parameters of said second artificial neural network in a first storage area, and storing parameters of said dummy nodes in a second storage area; and
while an inference is to be undertaken, by a host computer, respectively retrieving, by the host computer, said model parameters of said second artificial neural network and said parameters of said dummy nodes from said first storage area and said second storage area simultaneously; deleting interconnections between said dummy nodes of said second artificial neural network or setting said interconnections between said dummy nodes of said second artificial neural network to 0 according to said parameters of said dummy nodes before said inference,
wherein a method for establishing said dummy nodes includes steps:
inserting said dummy nodes into said first artificial neural network to vary an architecture of said first artificial neural network; and
generating a plurality of interconnections between said dummy nodes, and randomly assigning weights to said interconnections to form said second artificial neural network.