US 12,124,779 B2
Method of neural network construction for the simulation of physical systems
Manuel Bompard, Ramonville (FR); Mathieu Causse, Ramonville (FR); Florent Masmoudi, Ramonville (FR); Mohamed Masmoudi, Ramonville (FR); and Houcine Turki, Ramonville (FR)
Assigned to ADAGOS, Ramonville-Saint-Agne (FR)
Appl. No. 17/292,305
Filed by ADAGOS, Ramonville-Saint-Agne (FR)
PCT Filed Nov. 7, 2019, PCT No. PCT/FR2019/052649
§ 371(c)(1), (2) Date May 7, 2021,
PCT Pub. No. WO2020/094995, PCT Pub. Date May 14, 2020.
Claims priority of application No. 1860383 (FR), filed on Nov. 9, 2018; and application No. 1900572 (FR), filed on Jan. 23, 2019.
Prior Publication US 2021/0397770 A1, Dec. 23, 2021
Int. Cl. G06F 30/27 (2020.01); G06N 3/04 (2023.01); G06N 3/082 (2023.01)
CPC G06F 30/27 (2020.01) [G06N 3/04 (2013.01); G06N 3/082 (2013.01)] 16 Claims
OG exemplary drawing
 
1. A method of construction of a feedforward neural network, the method comprising:
initializing a neural network according to an initial network topology comprising an input layer, at least one hidden layer comprising at least one node, and a set of output nodes, each layer being defined by a set of simultaneously calculable nodes; and
implementing at least one topological optimization phase of the neural network, wherein each topological optimization phase comprises:
at least one additive phase, comprising a modification of the network topology by adding at least one node and/or a connection link between an input of a node of a layer and an output of a node of any preceding layer; and/or
at least one subtractive phase comprising a modification of the network topology by removing at least one node and/or a connection link between nodes of two layers,
wherein each modification of the network topology during an additive or subtractive phase comprises:
for a plurality of candidate modifications of the network topology, estimating a variation between:
a network error for a current topology, and
a network error for a candidate modified topology, the candidate modified topology being a topology resulting from a candidate modification applied to the current topology, and
selecting a modification of the network topology among the plurality of candidate modifications, based on the estimated variations of the network error.