US 11,720,795 B2
Neural network structure and a method thereto
Harri Valpola, Helsinki (FI)
Assigned to Canary Capital LLC, Wilmington, DE (US)
Appl. No. 15/531,212
Filed by CANARY CAPITAL LLC, Wilmington, DE (US)
PCT Filed Nov. 26, 2014, PCT No. PCT/FI2014/050911
§ 371(c)(1), (2) Date May 26, 2017,
PCT Pub. No. WO2016/083657, PCT Pub. Date Jun. 2, 2016.
Prior Publication US 2017/0330076 A1, Nov. 16, 2017
Int. Cl. G06N 3/084 (2023.01); G06N 20/00 (2019.01); G06N 3/02 (2006.01); G06N 7/04 (2006.01)
CPC G06N 3/084 (2013.01) [G06N 3/02 (2013.01); G06N 7/046 (2013.01); G06N 20/00 (2019.01)] 22 Claims
OG exemplary drawing
 
1. A system, comprising:
a memory; and
one or more processors configured to execute instructions stored in the memory to define a neural network structure that is trained using one or more cost functions and includes a base layer and a second layer, wherein:
the base layer includes a corruption function and a decoding function,
the second layer includes an encoding function and a decoding function,
the corruption function of the base layer receives input data of the neural network structure as an input and generates corrupted input data,
the encoding function of the second layer receives the corrupted input data from the corruption function of the base layer as an input and generates a second layer encoding output,
the decoding function of the second layer receives the second layer encoding output from the encoding function of the second layer as an input and generates a second layer decoding output,
the decoding function of the base layer receives the second layer decoding output from the decoding function of the second layer as an input, receives the corrupted input data from the corruption function of the base layer as a lateral input, and generates a base layer decoding output, and
the one or more cost functions include a cost function of the base layer, the cost function of the base layer receives the base layer decoding output from the decoding function of the base layer as an input and receives the input data of the neural network structure as an input.