US 12,443,891 B2
Vector processing of decision trees to form inferences
Jan Van Lunteren, Rüschlikon (CH)
Assigned to International Business Machines Corporation, Armonk, NY (US)
Filed by International Business Machines Corporation, Armonk, NY (US)
Filed on Aug. 25, 2021, as Appl. No. 17/445,888.
Prior Publication US 2023/0068120 A1, Mar. 2, 2023
Int. Cl. G06N 20/20 (2019.01); G06F 9/30 (2018.01); G06F 9/38 (2018.01)
CPC G06N 20/20 (2019.01) [G06F 9/30029 (2013.01); G06F 9/30036 (2013.01); G06F 9/30098 (2013.01); G06F 9/3887 (2013.01); G06F 9/3888 (2023.08)] 20 Claims
OG exemplary drawing
 
1. A computer-implemented method of performing machine learning inferences, comprising:
monitoring characteristics of a computerized system used to perform the machine learning inferences, the monitored characteristic comprising a number of available vector registers and a current load of the computerized system;
selecting a set of input records to be processed by decision trees on the computerized system;
identifying operations to be performed as matrix elements, wherein the matrix elements correspond, each, to one of the input records to be processed by one of the decision trees;
monitoring characteristics of one or each of the matrix elements;
dynamically partitioning the matrix elements, based on the monitored characteristics of the one or each of the matrix elements and the computerized system used to perform the machine learning inferences, to determine disjoint subsets of the matrix elements; and
concurrently processing, by the computerized system, the disjoint subsets of the matrix elements based on the dynamic partitioning, wherein the concurrently processing comprises:
processing a first disjoint subset of the disjoint subsets using vector processing, based on vector instructions operating on data stored in the vector registers, such that the matrix elements of the first disjoint subset are processed in parallel;
determining that at least one leaf node has been reached for at least one of the decision trees involved in the first disjoint subset;
determining whether a termination condition has been met, wherein the termination condition is met when a leaf node has been reached for all of the decision trees involved in the first disjoint subset; and
in response to determining that the termination condition has not been met, processing each of the at least one leaf node that has been reached as a split node looping to itself using a pointer until the termination condition is met.