US 11,900,052 B2
Automatic generation of transformations of formatted templates using deep learning modeling
Ji Li, San Jose, CA (US); Amit Srivastava, San Jose, CA (US); and Mingxi Cheng, Los Angeles, CA (US)
Assigned to MICROSOFT TECHNOLOGY LICENSING, LLC, Redmond, WA (US)
Filed by Microsoft Technology Licensing, LLC, Redmond, WA (US)
Filed on Nov. 11, 2020, as Appl. No. 17/095,603.
Prior Publication US 2022/0147702 A1, May 12, 2022
Int. Cl. G06N 3/02 (2006.01); G06N 3/088 (2023.01); G06N 5/04 (2023.01); G06F 40/186 (2020.01); G06N 20/00 (2019.01); G06F 40/103 (2020.01); G06V 10/40 (2022.01)
CPC G06F 40/186 (2020.01) [G06F 40/103 (2020.01); G06N 3/02 (2013.01); G06N 3/088 (2013.01); G06N 5/04 (2013.01); G06N 20/00 (2019.01); G06V 10/40 (2022.01)] 20 Claims
OG exemplary drawing
 
1. A method comprising:
extracting feature data for objects of a first slide-based template and feature data for visual style attributes of the first slide-based template, wherein the feature data for the objects comprises shape information for the objects, and wherein the first slide-based template is associated with a first presentation theme providing a first set of visual style attributes for the objects of the first slide-based template;
applying trained artificial intelligence (AI) processing, configured for generation of transformations of slide-based templates, to generate a transformation of the first slide-based template, wherein the applying of the trained AI processing executes processing operations that comprise:
encoding the feature data for the objects of the first slide-based template as a latent vector providing a distributed representation of the feature data,
propagating the latent vector to a first decoder network selected from a plurality of decoder networks, wherein each of the plurality of decoder networks is trained based on training data comprising data for slide-based templates having one of a plurality of different presentation themes, wherein the first decoder network is trained based on training data comprising data for slide-based templates having a second presentation theme of the plurality of different presentation themes, wherein the second presentation theme is different from the first presentation theme and provides a second set of visual style attributes for objects thereof, and
automatically generating a transformed template for the first slide-based template based on analysis of the latent vector using the first decoder network, wherein the transformed template comprises: one or more transformations of the objects of the first slide-based template and a style transformation modifying one or more visual style attributes of the first set of visual style attributes; and
storing the transformed template for subsequent presentation through a productivity service.