US 11,900,237 B2
Organizing neural networks
Henry Markram, Lausanne (CH); Rodrigo de Campos Perin, Lausanne (CH); and Thomas K. Berger, Berkeley, CA (US)
Assigned to ECOLE POLYTECHNIQUE FEDERALE DE LAUSANNE (EPFL), Lausanne (CH)
Filed by Ecole Polytechnique Federale De Lausanne (EPFL), Lausanne (CH)
Filed on Sep. 20, 2021, as Appl. No. 17/479,180.
Application 14/838,013 is a division of application No. 13/566,128, filed on Aug. 3, 2012, granted, now 10,387,767, issued on Aug. 20, 2019.
Application 17/479,180 is a continuation of application No. 16/528,807, filed on Aug. 1, 2019, granted, now 11,126,911, issued on Sep. 21, 2021.
Application 16/528,807 is a continuation of application No. 14/838,013, filed on Aug. 27, 2015, granted, now 10,373,048, issued on Aug. 6, 2019.
Application 13/566,128 is a continuation of application No. PCT/EP2011/000515, filed on Feb. 4, 2011.
Claims priority of provisional application 61/301,781, filed on Feb. 5, 2010.
Prior Publication US 2022/0121907 A1, Apr. 21, 2022
Prior Publication US 2023/0394280 A9, Dec. 7, 2023
Int. Cl. G06N 3/045 (2023.01); G06N 3/082 (2023.01); G06N 3/08 (2023.01)
CPC G06N 3/045 (2023.01) [G06N 3/08 (2013.01); G06N 3/082 (2013.01)] 11 Claims
OG exemplary drawing
 
1. A machine-implemented method of forming a neural network device, the method comprising:
forming a network of nodes implemented in hardware, in software, or in a combination thereof; and
assigning or reassigning links between nodes of the network by connecting or disconnecting nodes with a probability that embodies a number of common neighbors shared by the nodes, wherein the assigning or reassigning forms a plurality of node assemblies that are interconnected by between-assembly links, wherein each node assembly comprises a network of nodes interconnected by a plurality of within-assembly links, wherein the assigning or reassigning is repeated until the number of links within each respective node assembly exceeds the number of nodes within that node assembly, wherein a number of reciprocal links is lower than a number of non-reciprocal links; and
training the network of nodes, wherein training the network of nodes comprises
weighting within-assembly links with weights that do not embody the training, and
adapting weights of the between-assembly links to embody the training.