Granted Patent B2

US 12,430,582 · App. 17/376,962 · Granted Sep 30, 2025

Inventors: Andrew Feng (Cupertino, CA); Jun Shi (Sunnyvale, CA); Mridul Jain (Cupertino, CA); Peter Cnudde (Los Altos, CA)

Assignee: Verizon Patent and Licensing Inc.

G06N20/00

Abstract

Claims (61)

1. A method implemented on at least one computing device each of which has at least one processor, storage, and a communication platform connected to a network for estimating one or more parameters of a machine learning model, the method comprising:

estimating, by each of a plurality of processing units in a first node of a plurality of nodes, values of the one or more parameters associated with the machine learning model, wherein each of the plurality of processing units estimates the values based on a batch of a subset of training data allocated to the first node;

generating, by a preselected representative processing unit of the first node preselected from the plurality of the processing units, a vector by aggregating the values of the one or more parameters estimated by the plurality of processing units in the first node;

dividing, by the representative processing unit of the first node, the vector into a plurality of fractions, each of which is designated to one of the plurality of nodes;

identifying a first fraction and remaining fractions of the plurality of fractions, wherein the first fraction is designated to the first node and the remaining fractions are designated to corresponding other nodes of the plurality of nodes;

aggregating the first fraction designated to the first node and the first fractions received from other nodes of the plurality of nodes to generate an aggregated first fraction;

sending, by the representative processing unit of the first node, the aggregated first fraction to representative processing units of the other nodes of the plurality of nodes;

receiving, by the representative processing unit of the first node, from the representative processing units of the other nodes of the plurality of nodes, aggregated remaining fractions of the remaining fractions designated to the other nodes;

creating an iteration final estimate of the one or more parameters based on the aggregated first fraction and the aggregated remaining fractions;

broadcasting, by the representative processing unit of the first node, the iteration final estimate to all other processing units of the plurality of processing units in the first node to synchronize a learning state of the machine-learning model across the plurality of nodes; and

iterating the above recited steps until a predetermined condition is met.

2. The method of claim 1 , further comprising:

generating a snapshot of a state of the one or more parameters of the model; and

storing the snapshot in a storage outside the plurality of nodes.

3. The method of claim 2 , further comprising:

in response to a failure with respect to estimating the one or more parameters, retrieving the snapshot of the state of the one or more parameters of the model from the storage; and

continuing to estimate the one or more parameters based on the last state of the one or more parameters.

4. The method claim 1 , wherein the plurality of nodes are synchronized via a Message Passing Interface (MPI) AllReduce based Application Program Interface (API) using a network interface implemented on each of the plurality of nodes, wherein the network interface includes an Ethernet interface, an Infiniband interface, or the Ethernet interface and the Infiniband interface.

5. The method of claim 1 , wherein the preselected representative processing unit is a graphical processing unit.

6. A system, having at least one processor, storage, and a communication platform connected to a network for estimating one or more parameters of a machine learning model, the system comprising:

a first node, of a plurality of nodes, implemented by the at least one processor and configured for:

estimating, by each of a plurality of processing units in the first node, values of the one or more parameters associated with the machine learning model, wherein each of the plurality of processing units estimates the values based on a batch of a subset of training data allocated to the first node;

generating, by a preselected representative processing unit of the first node preselected from the plurality of the processing units, a vector by aggregating the values of the one or more parameters estimated by the plurality of processing units in the first node;

dividing, by the representative processing unit of the first node, the vector into a plurality of fractions, each of which is designated to one of the plurality of nodes;

identifying a first fraction and remaining fractions of the plurality of fractions, wherein the first fraction is designated to the first node and the remaining fractions are designated to corresponding other nodes of the plurality of nodes;

aggregating the first fraction designated to the first node and the first fractions received from other nodes of the plurality of nodes to generate an aggregated first fraction;

sending, by the representative processing unit of the first node, the aggregated first fraction to representative processing units of the other nodes of the plurality of nodes;

receiving, by the representative processing unit of the first node, from the representative processing units of the other nodes of the plurality of nodes, aggregated remaining fractions of the remaining fractions designated to the other nodes;

creating an iteration final estimate of the one or more parameters based on the aggregated first fraction and the aggregated remaining fractions;

broadcasting, by the representative processing unit of the first node, the iteration final estimate to all other processing units of the plurality of processing units in the first node to synchronize a learning state of the machine-learning model across the plurality of nodes; and

iterating the above recited steps until a predetermined condition is met; and

the other nodes, of the plurality of nodes, implemented by the at least one processor and configured for:

receiving the aggregated first fraction of the vector from the first node; and

sending the aggregated remaining fractions of the vector designated to the other nodes to the first node.

7. The system of claim 6 , wherein the first node comprises:

a model snapshot generator implemented by the at least one processor and configured for

generating a snapshot of a state of the one or more parameters of the model; and

storing the snapshot in a storage outside the plurality of nodes.

8. The system of claim 7 , wherein the first node is implemented by the at least one processor and configured further for:

in response to a failure with respect to estimating the one or more parameters, retrieving the snapshot of the state of the one or more parameters of the model from the storage; and

continuing to estimate the one or more parameters based on the last state of the one or more parameters.

9. The system claim 6 , wherein the plurality of nodes are synchronized via a Message Passing Interface (MPI) AllReduce based Application Program Interface (API) using a network interface implemented on each of the plurality of nodes, wherein the network interface includes an Ethernet interface, an Infiniband interface, or the Ethernet interface and the Infiniband interface.

10. The system of claim 6 , wherein the preselected representative processing unit is a graphical processing unit.

11. A machine-readable tangible and non-transitory medium having information for estimating one or more parameters of a machine learning model, wherein the information, when read by a first node of a plurality of nodes, causes the first node to perform the following:

estimating, by each of a plurality of processing units in the first node of the plurality of nodes, values of the one or more parameters associated with the machine learning model, wherein each of the plurality of processing units estimates the values based on a batch of a subset of training data allocated to the first node;

generating, by a preselected representative processing unit of the first node preselected from the plurality of the processing units, a vector by aggregating the values of the one or more parameters estimated by the plurality of processing units in the first node;

dividing, by the representative processing unit of the first node, the vector into a plurality of fractions, each of which is designated to one of the plurality of nodes;

identifying a first fraction and remaining fractions of the plurality of fractions, wherein the first fraction is designated to the first node and the remaining fractions are designated to corresponding other nodes of the plurality of nodes;

aggregating the first fraction designated to the first node and the first fractions received from other nodes of the plurality of nodes to generate an aggregated first fraction;

sending, by the representative processing unit of the first node, the aggregated first fraction to representative processing units of the other nodes of the plurality of nodes;

receiving, by the representative processing unit of the first node, from the representative processing units of the other nodes of the plurality of nodes, aggregated remaining fractions of the remaining fractions designated to the other nodes;

creating an iteration final estimate of the one or more parameters based on the aggregated first fraction and the aggregated remaining fractions;

broadcasting, by the representative processing unit of the first node, the iteration final estimate to all other processing units of the plurality of processing units in the first node to synchronize a learning state of the machine-learning model across the plurality of nodes; and

iterating the above recited steps until a predetermined condition is met.

12. The machine-readable tangible and non-transitory medium of claim 11 , wherein the information, when read by the first node, causes the first node to further perform:

generating a snapshot of a state of the one or more parameters of the model; and

storing the snapshot in a storage outside the plurality of nodes.

13. The machine-readable tangible and non-transitory medium of claim 12 , wherein the information, when read by the first node, causes the first node to further perform:

in response to a failure with respect to estimating the one or more parameters, retrieving the snapshot of the state of the one or more parameters of the model from the storage; and

continuing to estimate the one or more parameters based on the last state of the one or more parameters.

14. The machine-readable tangible and non-transitory medium of claim 11 , wherein the plurality of nodes are synchronized via a Message Passing Interface (MPI) AllReduce based Application Program Interface (API) using a network interface implemented on each of the plurality of nodes, wherein the network interface includes an Ethernet interface, an Infiniband interface, or the Ethernet interface and the Infiniband interface.

Assignments (6)

Continuity (2)

References Cited (26)