US 11,915,120 B2
Flexible parameter sharing for multi-task learning
Effrosyni Kokiopoulou, Horgen (CH); Krzysztof Stanislaw Maziarz, Cambridge (GB); Andrea Gesmundo, Zurich (CH); Luciano Sbaiz, Gattikon (CH); Gábor Bartók, Zurich (CH); and Jesse Berent, Geneva (CH)
Assigned to GOOGLE LLC, Mountain View, CA (US)
Filed by Google LLC, Mountain View, CA (US)
Filed on Mar. 17, 2020, as Appl. No. 16/820,829.
Claims priority of application No. 20200100034 (GR), filed on Jan. 27, 2020.
Prior Publication US 2021/0232895 A1, Jul. 29, 2021
Int. Cl. G06N 3/044 (2023.01); G06N 20/00 (2019.01); G06N 3/084 (2023.01)
CPC G06N 3/044 (2023.01) [G06N 3/084 (2013.01); G06N 20/00 (2019.01)] 20 Claims
OG exemplary drawing
 
1. A computer-implemented method for training a machine-learned model for flexible-multi-task learning, the method comprising:
obtaining a test input;
selecting a particular task from a plurality of tasks; and
training the machine-learned model for the particular task, wherein training the machine-learned model for the particular task comprises:
obtaining a routing matrix associated with the particular task, the routing matrix comprising a plurality of values configured to activate one or more components of a layer of the machine-learned model based on the particular task, wherein a respective value of the routing matrix is obtained by:
sampling, using a respective probability value associated with a respective component and the particular task, the respective value of the routing matrix;
processing the test input using the machine-learned model to generate an output;
training the activated one or more components based at least in part on the output; and
training, using an approximation, the machine-learned model to obtain the routing matrix.