Granted Patent B2

US 12,547,875 · App. 17/072,709 · Granted Feb 10, 2026

Inventor: Bradley David Safnuk (Victoria, CA)

Assignee: Ford Global Technologies, LLC

G06N3/045G06N3/08

Abstract

Claims (54)

1 . A method of distributed computing, the method comprising:

dividing a directed acyclic graph (DAG) among a plurality of nodes, the directed acyclic graph comprising a plurality of vertices linked in pairwise relationships via a plurality of edges, wherein each of the plurality of nodes comprises a computing device;

performing a topological sort of the vertices of the directed acyclic graph, the topological sort comprising a deterministic topological sort controlling an order of operations for traversing the directed acyclic graph;

generating a DAG map of the directed acyclic graph that comprises the deterministic topological sort and that comprises an indication of the order of operations for traversing the directed acyclic graph among the plurality of nodes;

providing the DAG map of the directed acyclic graph to each vertex of the plurality of vertices of the directed acyclic graph such that each vertex of the plurality of vertices receives (i) a description of a flow of data through the plurality of vertices represented by the deterministic topological sort and (ii) a processing order of traversal of the plurality of vertices of the directed acyclic graph, wherein when a particular vertex of the plurality of vertices receives the DAG map, the particular vertex prepares to execute an operation to be performed by the particular vertex as indicated by the DAG map prior to an actual execution of the operation by the particular vertex;

training a neural network associated with a vehicle system by using a distributed computing system to traverse the directed acyclic graph; and

executing the trained neural network to facilitate a human machine interface of the vehicle system by generating one or more predictions associated with the vehicle system to simulate a behavior or an attribute of the vehicle system.

2 . The method of claim 1 , wherein the directed acyclic graph forms a part of the neural network.

3 . The method of claim 2 , further comprising: identifying an edge linking separate nodes; and inserting a data exchange vertex between the separate nodes.

4 . The method of claim 3 , wherein identifying the edge linking separate nodes comprises: tracing an edge from a first vertex to a second vertex, wherein the edge coupled the first vertex to the second vertex; comparing a node of the first vertex to a node of the second vertex; and determining that the edge links separate nodes when the node of the first vertex is different than the node of the second vertex.

5 . The method of claim 4 , wherein the data exchange vertex comprises a recursive directed acyclic graph structure comprising:

a send vertex located on the node of the first vertex; and

a receive vertex located on the node of the second vertex, wherein when the data exchange vertex is activated, both the send vertex and the receive vertex run in concert.

6 . The method of claim 5 , further comprising:

activating a data vertex; and

running the send vertex and the receive vertex in concert while passing a tensor from the send vertex to the receive vertex.

7 . The method of claim 6 , wherein passing the tensor from the send vertex to the receive vertex comprises:

passing a series of communications characterizing the tensor from the send vertex to the receive vertex; and

subsequently passing the tensor from the send vertex to the receive vertex.

8 . The method of claim 7 , wherein passing the series of communications characterizing the tensor from the send vertex to the receive vertex comprises: passing a first value indicating a rank of the tensor from the send vertex to the receive vertex; passing a vector of values indicating a shape of the tensor from the send vertex to the receive vertex; passing at least one second value characterizing data contained in the tensor from the send vertex to the receive vertex; and passing a third value directing tracking of a gradient of the Neural Network.

9 . The method of claim 1 , further comprising:

creating at least one clone directed acyclic graph identical to the directed acyclic graph, the clone directed acyclic graph comprising a plurality of clone vertices;

for each of the clone vertices, identifying a corresponding vertex in the directed acyclic graph;

during training of the directed acyclic graph and the clone directed acyclic graph, calculating aggregate gradient data based on gradient data from each of the clone vertices and its corresponding vertex in the directed acyclic graph; and

updating at least one weight of the directed acyclic graph and the clone directed acyclic graph based on the aggregate gradient data.

10 . The method of claim 1 , wherein one of the plurality of vertices of the directed acyclic graph comprises an entry vertex, the method further comprising:

identifying the nodes underlying the directed acyclic graph;

generating a subordinate directed acyclic graph in the entry vertex, the subordinate directed acyclic graph comprising a plurality of subordinate vertices, each of the plurality of subordinate vertices corresponding to a one of the nodes underlying the directed acyclic graph;

receiving data and metadata at the entry vertex;

delivering the data to a next vertex in the directed acyclic graph; and

communicating the metadata to nodes underlying the directed acyclic graph via the subordinate directed acyclic graph.

11 . A system comprising:

one or more processors; and

one or more memories storing computer-executable instructions that, when executed by the one or more processors, configure the one or more processors to:

divide a directed acyclic graph among a plurality of nodes, the directed acyclic graph comprising a plurality of vertices linked in pairwise relationships via a plurality of edges, wherein each of the plurality of nodes comprises a computing device;

perform a topological sort of the vertices of the directed acyclic graph, the topological sort comprising a deterministic topological sort controlling an order of operations for traversing the directed acyclic graph;

generate a DAG map of the directed acyclic graph that comprises the deterministic topological sort and that comprises an indication of the order of operations for traversing the directed acyclic graph among the plurality of nodes;

provide the DAG map of the directed acyclic graph to each vertex of the plurality of vertices of the directed acyclic graph such that each vertex of the plurality of vertices receives (i) a description of a flow of data through the plurality of vertices represented by the deterministic topological sort and (ii) a processing order of traversal of the plurality of vertices of the directed acyclic graph, wherein when a particular vertex of the plurality of vertices receives the DAG map, the particular vertex prepares to execute an operation to be performed by the particular vertex as indicated by the DAG map prior to an actual execution of the operation by the particular vertex;

train a neural network associated with a vehicle system by traversing the directed acyclic graph; and

execute the trained neural network to facilitate a human machine interface of the vehicle system by generating one or more predictions associated with the vehicle system to simulate a behavior or an attribute of the vehicle system.

12 . The system of claim 11 , wherein the directed acyclic graph forms a part of the neural network.

13 . The system of claim 12 , further comprising: identifying an edge linking separate nodes; and inserting a data exchange vertex between the separate nodes.

14 . The system of claim 13 , wherein identifying the edge linking separate nodes comprises: tracing an edge from a first vertex to a second vertex, wherein the edge coupled the first vertex to the second vertex; comparing a node of the first vertex to a node of the second vertex; and determining that the edge links separate nodes when the node of the first vertex is different than the node of the second vertex.

15 . The system of claim 14 , wherein the data exchange vertex comprises:

a recursive directed acyclic graph structure;

a send vertex located on the node of the first vertex; and

a receive vertex located on the node of the second vertex, wherein when the data exchange vertex is activated, both the send vertex and the receive vertex run in concert.

16 . The system of claim 15 , further comprising:

activating a data vertex; and

running the send vertex and the receive vertex in concert while passing a tensor from the send vertex to the receive vertex.

17 . The system of claim 16 , wherein passing the tensor from the send vertex to the receive vertex comprises:

passing a series of communications characterizing the tensor from the send vertex to the receive vertex; and

subsequently passing the tensor from the send vertex to the receive vertex.

18 . The system of claim 17 , wherein passing the series of communications characterizing the tensor from the send vertex to the receive vertex comprises: passing a first value indicating a rank of the tensor from the send vertex to the receive vertex; passing a vector of values indicating a shape of the tensor from the send vertex to the receive vertex; passing at least one second value characterizing data contained in the tensor from the send vertex to the receive vertex; and passing a third value directing tracking of a gradient of the Neural Network.

Assignments (1)

Continuity (1)

References Cited (38)