US 11,948,067 B1
Storing of intermediate computed values for subsequent use in a machine trained network
Ryan J. Cassidy, San Diego, CA (US); and Steven L. Teig, Menlo Park, CA (US)
Assigned to PERCEIVE CORPORATION, San Jose, CA (US)
Filed by Perceive Corporation, San Jose, CA (US)
Filed on Nov. 9, 2020, as Appl. No. 17/093,278.
Claims priority of provisional application 62/933,960, filed on Nov. 11, 2019.
This patent is subject to a terminal disclaimer.
Int. Cl. G06N 3/04 (2023.01); G06N 3/049 (2023.01); G06N 3/063 (2023.01)
CPC G06N 3/049 (2013.01) [G06N 3/063 (2013.01)] 26 Claims
OG exemplary drawing
 
1. A non-transitory machine readable medium storing a program for implementing a temporal convolution network (TCN) comprising a plurality of layers of machine-trained (MT) processing nodes, the program comprising sets of instructions for:
propagating first set of input values through the layers of the TCN to compute a first output of the TCN for the first set of input values, the first set of input values received by the TCN at a first instance in time, the propagation of the first set of input values comprising computation of a first plurality of activation values;
storing the first plurality of activation values in a non-transitory machine-readable storage for re-use during propagation of other sets of input values through the TCN;
during propagation through the layers of the TCN of a second set of input values received by the TCN at a second instance in time, retrieving the first plurality of activation values from the non-transitory machine-readable storage and using the retrieved first plurality of activation values to compute a second plurality of activation values in order to compute a second output of the TCN for the second set of input values; and
during propagation through the layers of the TCN of a third set of input values received by the TCN at a third instance in time, retrieving the first plurality of activation values from the non-transitory machine-readable storage and using the retrieved first plurality of activation values to compute a third plurality of activation values in order to compute a third output of the TCN for the third set of input values,
wherein the first, second, and third pluralities of activation values are computed respectively by first, second, and third sets of MT processing nodes of the TCN,
wherein the second and third sets of MT processing nodes are different sets of MT processing nodes of the TCN such that the first plurality of activation values are used by different sets of MT processing nodes when computing the second output of the TCN for the second set of input values than when computing the third output of the TCN for the third set of input values.