US 12,367,384 B2
Residual semi-recurrent neural networks
Qi Tang, Cambridge, MA (US); and Youran Qi, Madison, WI (US)
Assigned to Sanofi, Paris (FR)
Filed by Sanofi, Paris (FR)
Filed on Apr. 4, 2024, as Appl. No. 18/626,358.
Application 18/626,358 is a continuation of application No. 18/179,890, filed on Mar. 7, 2023, granted, now 11,977,972.
Application 18/179,890 is a continuation of application No. 16/827,094, filed on Mar. 23, 2020, granted, now 11,625,589, issued on Apr. 11, 2023.
Claims priority of provisional application 62/824,895, filed on Mar. 27, 2019.
Claims priority of application No. 19305611 (EP), filed on May 13, 2019.
Prior Publication US 2024/0265246 A1, Aug. 8, 2024
Int. Cl. G06N 3/063 (2023.01); G01N 33/50 (2006.01); G06N 3/045 (2023.01); G06N 3/082 (2023.01)
CPC G06N 3/063 (2013.01) [G01N 33/5008 (2013.01); G06N 3/045 (2023.01); G06N 3/082 (2013.01)] 25 Claims
OG exemplary drawing
 
1. A method performed by one or more computers, the method comprising:
receiving data characterizing a subject, the data comprising:
(i) a sequence of time-varying features of the subject, wherein the sequence of time-varying features includes a respective time-varying feature of the subject for each time point in a sequence of time points; and
(ii) a set of time-invariant features of the subject, wherein the time-invariant features of the subject are the same for each time point in the sequence of time points;
processing the set of time-invariant features of the subject using a multilayer perceptron to generate an encoded representation of the set of time-invariant features of the subject;
processing the sequence of time-varying features of the subject using a recurrent neural network to generate an encoded representation of the sequence of time-varying features of the subject, comprising, for each time point in the sequence of time points:
processing the time-varying feature of the subject for the time point and a hidden state of the recurrent neural network to update the hidden state of the recurrent neural network;
generating a combined feature representation by combining: (i) the encoded representation of the set of time-invariant features of the subject generated by the multilayer perceptron, and (ii) the encoded representation of the set of time-varying features of the subject generated by the recurrent neural network; and
processing the combined feature representation to generate one or more predictions characterizing the subject.