Systems and methods for graph-based AI training
Graphs are powerful structures made of nodes and edges. Information can be encoded in the nodes and edges themselves, as well as the connections between them. Graphs can be used to create manifolds which in turn can be used to efficiently train more robust AI systems. Systems and methods for graph-based AI training in accordance with embodiments of the invention are illustrated. In one embodiment, a graph interface system including a processor, and a memory configured to store a graph interface application, where the graph interface application directs the processor to obtain a set of training data, where the set of training data describes a plurality of scenarios, encode the set of training data into a first knowledge graph, generate a manifold based on the first knowledge graph, and train an AI model by traversing the manifold.
1. A graph interface system comprising:
a processor; and
a memory configured to store a graph interface application, where the graph interface application directs the processor to:
obtain a set of training data, where the set of training data describes a plurality of scenarios;
encode the set of training data into a first knowledge graph as a plurality of nodes representing spatiotemporal features of each scenario in the plurality of scenarios;
calculate a similarity measure between each scenario in the first knowledge graph based on their spatiotemporal features; and
embed the scenarios into a manifold linked to the first knowledge graph by connecting scenarios via edges with an edge weight equal to the similarity measure between a first scenario and a second scenario connected by their respective edge; and
train an artificial intelligence (AI) model by traversing the manifold.
2. The graph interface system of claim 1 , wherein to encode the set of training data into the knowledge graph, the graph interface application directs the processor to:
identify objects within each scenario in the plurality of scenarios;
store the identified objects within the first knowledge graph;
determine spatiotemporal features for each object;
store the spatiotemporal features within the first knowledge graph; and
generate spatiotemporal scenarios by clustering the first knowledge graph based on the spatiotemporal features.
3. The graph interface system of claim 2 , wherein the graph interface application further directs the processor to add semantic meaning to spatiotemporal features.
4. The graph interface system of claim 3 , wherein to add semantic meaning to spatiotemporal features, the graph interface application further directs the processor to:
obtain a second knowledge graph from a kernel set of annotated spatiotemporal scenarios;
merge the first knowledge graph and the second knowledge graph;
assign annotations from scenarios in the second knowledge subgraph to identical scenarios in the first knowledge subgraph;
obtain a third knowledge subgraph encoding only descriptions of spatiotemporal scenarios;
merge the third knowledge subgraph into the first knowledge subgraph;
assign spatiotemporal scenarios in the first knowledge graph to descriptions from the third knowledge subgraph using annotations from the second knowledge subgraph; and
assign still-unconnected descriptions from the third graph to still-unannotated scenarios in the first knowledge graph via an inference method.
5. The graph interface system of claim 4 , wherein the inference method is a Bayesian inference method.
6. The graph interface system of claim 1 , wherein to generate a manifold based on the first knowledge graph, the graph interface application further directs the processor to prune illegal spatiotemporal scenarios.
7. The graph interface system of claim 1 , wherein to traverse the manifold, the graph interface application directs the processor to:
begin at a centermost node in the manifold; and
successively navigate to a new node that is the greatest total distance from all previously visited nodes.
8. The graph interface system of claim 1 , wherein the AI model is used in an autonomous vehicle control system.
9. The graph interface system of claim 1 , wherein the training data comprises forward-mounted monocular camera footage and textual accident reports.
10. The graph interface system of claim 1 , wherein the first knowledge graph is constructed using dimensional data structures.
11. The graph interface system of claim 1 , wherein the graph interface application further directs the processor to audit at least one scenario in the manifold.
12. The graph interface system of claim 1 , wherein the graph interface application further directs the processor to provide a projection of the first knowledge graph using a hybrid interface.
13. The system of claim 1 , wherein the graph interface application directs the processor to:
encode a risk value reflective of the scenario represented by a given node; and
train the AI by providing the AI with nodes in the manifold based on highest risk, where each successive provided node is from a different cluster of nodes in the manifold.
14. A method for training an AI using a graph interface system, comprising:
obtaining a set of training data, where the set of training data describes a plurality of spatiotemporal scenarios;
encoding the set of training data into a first knowledge graph as a plurality of nodes representing spatiotemporal features of each scenario in the plurality of scenarios;
calculating a similarity measure between each scenario in the first knowledge graph based on their spatiotemporal features; and
embedding the scenarios into a manifold linked to the first knowledge graph by connecting scenarios via edges with an edge weight equal to the similarity measure between a first scenario and a second scenario connected by their respective edge; and
training an artificial intelligence (AI) model by traversing the manifold.
15. The method of claim 14 , wherein encoding the set of training data into the knowledge graph comprises:
identifying objects within each scenario in the plurality of scenarios;
storing the identified objects within the first knowledge graph;
determining spatiotemporal features for each object;
storing the spatiotemporal features within the first knowledge graph; and
generating spatiotemporal scenarios by clustering the first knowledge graph based on the spatiotemporal features.
16. The method of claim 15 , further comprising adding semantic meaning to spatiotemporal features.
17. The method of claim 16 , adding semantic meaning to spatiotemporal features comprises:
obtaining a second knowledge graph from a kernel set of annotated spatiotemporal scenarios;
merging the first knowledge graph and the second knowledge graph;
assigning annotations from scenarios in the second knowledge subgraph to identical scenarios in the first knowledge subgraph;
obtaining a third knowledge subgraph encoding only descriptions of spatiotemporal scenarios;
merging the third knowledge subgraph into the first knowledge subgraph;
assign spatiotemporal scenarios in the first knowledge graph to descriptions from the third knowledge subgraph using annotations from the second knowledge subgraph; and
assigning still-unconnected descriptions from the third graph to still-unannotated scenarios in the first knowledge graph via an inference method.
18. The method of claim 17 , wherein the inference method is a Bayesian inference method.
19. The method of claim 14 , generating a manifold based on the first knowledge graph comprises pruning illegal spatiotemporal scenarios.
20. The method of claim 14 , further comprising:
encoding a risk value reflective of the scenario represented by a given node; and
training the AI by providing the AI with nodes in the manifold based on highest risk, where each successive provided node is from a different cluster of nodes in the manifold.