US 11,720,797 B2
Differential recurrent neural network
Patrice Simard, Bellevue, WA (US)
Assigned to Microsoft Technology Licensing, LLC, Redmond, WA (US)
Filed by Microsoft Technology Licensing, LLC, Redmond, WA (US)
Filed on Apr. 28, 2020, as Appl. No. 16/860,555.
Application 16/860,555 is a continuation of application No. 15/488,221, filed on Apr. 14, 2017, granted, now 10,671,908.
Claims priority of provisional application 62/426,153, filed on Nov. 23, 2016.
Prior Publication US 2020/0327395 A1, Oct. 15, 2020
This patent is subject to a terminal disclaimer.
Int. Cl. G06N 3/084 (2023.01); G06V 10/82 (2022.01); G06N 3/044 (2023.01); G06V 10/764 (2022.01); G06F 18/24 (2023.01)
CPC G06N 3/084 (2013.01) [G06F 18/24 (2023.01); G06N 3/044 (2023.01); G06V 10/764 (2022.01); G06V 10/82 (2022.01)] 18 Claims
OG exemplary drawing
 
1. A method for providing a response from a differential recurrent neural network (RNN) based on elements included in an input vector, the method comprising:
receiving the input vector at a transition component of the differential RNN;
providing, by the transition component, an output to a differential non-linearity component, the output including a positive contribution vector and a negative contribution vector corresponding to elements of states stored at a state component;
generating, by the differential non-linearity component, a state contribution vector based on the positive contribution vector and the negative contribution vector;
receiving, at the state component, the state contribution vector; and
providing as the response, an output based on stored state information received from the state component.