US 11,853,860 B2
Encoding and reconstructing inputs using neural networks
Martin Abadi, Palo Alto, CA (US); and David Godbe Andersen, Pittsburgh, PA (US)
Assigned to Google LLC, Mountain View, CA (US)
Filed by Google LLC, Mountain View, CA (US)
Filed on Mar. 3, 2022, as Appl. No. 17/685,559.
Application 17/685,559 is a continuation of application No. 16/323,205, granted, now 11,308,385, previously published as PCT/US2017/045329, filed on Aug. 3, 2017.
Claims priority of provisional application 62/370,954, filed on Aug. 4, 2016.
Prior Publication US 2023/0019228 A1, Jan. 19, 2023
This patent is subject to a terminal disclaimer.
Int. Cl. G06N 3/045 (2023.01); G06N 3/084 (2023.01)
CPC G06N 3/045 (2023.01) [G06N 3/084 (2013.01)] 15 Claims
OG exemplary drawing
 
1. A computer-implemented method for training a neural network system on one or more computers, comprising:
obtaining a plurality of training data pairs for the neural network system, each training data pair comprising a respective first input and a respective second input; and
training the neural network system on the plurality of training data pairs, comprising, for each training data pair:
processing, using an encoder neural network in accordance with current values of parameters of the encoder neural network, (i) the respective first input for the training data pair and (ii) the respective second input for the training data pair to generate an encoded representation of the respective first input for the training data pair, wherein the respective second input for the training data pair represents a symmetric encryption key;
processing, using a first decoder neural network in accordance with current values of parameters of the first decoder neural network, (i) the encoded representation of the respective first input for the training data pair and (ii) the respective second input for the training data pair to generate a first estimated reconstruction of the respective first input for the training data pair;
obtaining a second estimated reconstruction of the respective first input for the training data pair, the second estimated reconstruction generated by a second decoder neural network in accordance with current parameters of the second decoder neural network, by processing the encoded representation of the respective first input for the training data pair without processing the respective second input for the training data pair; and
adjusting the current parameters of the encoder neural network and the current parameters of the first decoder neural network based on the respective first input, the first estimated reconstruction of the respective first input, and the second estimated reconstruction of the respective first input.