Attention-based explanations for artificial intelligence behavior

1. A device configured to explain a result output by an artificial intelligence model configured to analyze an image, the device comprising:

a memory configured to store the artificial intelligence model and the image; and

a computation engine executing one or more processors,

wherein the computation engine is configured to receive a query regarding the image, and

wherein the computation engine is configured to execute the artificial intelligence model to analyze the image in order to output the result to the query,

wherein the artificial intelligence model is configured to, when analyzing the image to output the result:

segment the image into hierarchically arranged semantic areas in which one or more objects in the image are segmented into one or more parts;

determine, based on the query, an attention mask for the hierarchically arranged semantic areas in which an attention score is assigned to one or more of each of the one or more objects and each of the one or more parts;

update, based on the attention mask, the image to visually identify which of the hierarchically arranged semantic areas formed a basis for the result output by the artificial intelligence model; and

output the updated image.

2. The device of claim 1 , wherein the artificial intelligence model is configured to:

identify the one or more objects at a pixel-level of granularity;

segment the one or more objects into one or more parts at the pixel-level of granularity; and

determine the attention masks for the hierarchically arranged semantic areas at a pixel-level of granularity.

3. The device of claim 1 , wherein the artificial intelligence model is configured to recursively segment the image into the hierarchically arranged semantic areas.

4. The device of claim 1 , wherein the artificial intelligence model is configured to:

generate, based on the attention mask, a heatmap that uses color to visually identify which of the hierarchically arranged semantic areas formed a basis for the result output by the artificial intelligence model; and

overlay the heatmap over the image to obtain the updated image.

5. The device of claim 1 ,

wherein the artificial intelligence model includes a neural network configured to perform semantic parsing of the query to identify a meaning of the query, and

wherein the artificial intelligence model is further configured to analyze, based on the meaning, the hierarchically arranged semantic areas to obtain the result to the query.

6. The device of claim 1 , wherein the artificial intelligence model comprises a neural network configured to segment the image into the hierarchically arranged semantic areas.

7. The device of claim 1 , wherein the artificial intelligence model is configured to:

identify the one or more objects, the one or more parts, relationships between the one or more objects and the one or more parts, attributes associated with the one or more objects, and one or more scenes present in the image; and

generate a scene graph representing the one or more scenes as a hierarchical arrangement in which the one or more objects are associated with the corresponding one or more parts according to the relationships between the one or more objects and the one or more parts in which the one or more objects are associated with the corresponding attributes and one or more objects are associated with one more scenes.

8. The device of claim 1 ,

wherein the artificial intelligence model is further configured to:

receive an indication identifying one of the hierarchically arranged semantic areas; and

obtain, based on the attention mask, the attention score for the identified one of the hierarchically arranged semantic areas, and

wherein the artificial intelligence model is configured to update, based on the attention score for the identified one of the hierarchically arranged semantic areas, the identified one of the hierarchically arranged semantic areas of the image to visually identify an attention of the artificial intelligence model to the identified one of the hierarchically arranged semantic areas.

9. The device of claim 1 , wherein the artificial intelligence model is further configured to:

receive an indication identifying one of the hierarchically arranged semantic areas; and

modify the attention mask to change the attention score for the identified one of the hierarchically arranged areas.

10. A method of explaining a result output by an artificial intelligence model configured to analyze an image, the method comprising:

receiving, by a computing device, a query regarding the image; and

executing, by the computing device, the artificial intelligence model to analyze the image in order to output the result to the query,

wherein analyzing, by the artificial intelligence model, the image to output the result further comprises:

segmenting, by the computing device, the image into hierarchically arranged semantic areas in which one or more objects in the image are segmented into one or more parts;

determining, based on the query and by the computing device, an attention mask for the hierarchically arranged semantic areas in which an attention score is assigned to one or more of each of the one or more objects and each of the one or more parts;

updating, based on the attention mask and by the computing device, the image to visually identify which of the hierarchically arranged semantic areas formed a basis for the result output by the artificial intelligence model; and

outputting, by the computing device, the updated image.

11. The method of claim 10 , wherein segmenting the image comprises:

identifying the one or more objects at a pixel-level of granularity;

segmenting the one or more objects into one or more parts at the pixel-level of granularity; and

determining the attention masks for the hierarchically arranged semantic areas at a pixel-level of granularity.

12. The method of claim 10 , wherein segmenting the image comprises recursively segment the image into the hierarchically arranged semantic areas.

13. The method of claim 10 , further comprising:

generating, based on the attention mask, a heatmap that uses color to visually identify which of the hierarchically arranged semantic areas formed a basis for the result output by the artificial intelligence model; and

overlaying the heatmap over the image to obtain the updated image.

14. The method of claim 10 ,

wherein the artificial intelligence model includes a long short-term memory network configured to perform semantic parsing of the query to identify a meaning of the query, and

wherein the method further comprises analyzing, based on the meaning, the hierarchically arranged semantic areas to obtain the result to the query.

15. The method of claim 10 , wherein the artificial intelligence model comprises a neural network configured to segment the image into the hierarchically arranged semantic areas.

16. The method of claim 10 , wherein segmenting the image comprises:

identifying the one or more objects, the one or more parts, relationships between the one or more objects and the one or more parts, attributes associated with the one or more objects, and one or more scenes present in the image; and

generating a scene graph representing the one or more scenes as a hierarchical arrangement in which the one or more objects are associated with the corresponding one or more parts according to the relationships between the one or more objects and the one or more parts in which the one or more objects are associated with the corresponding attributes.

17. The method of claim 10 , wherein segmenting the image comprises segmenting the image into hierarchically arranged semantic areas in which the one or more objects in the image are segmented into the one or more parts, and the one or more parts are segmented into one or more sub-parts.

18. The method of claim 10 , further comprising:

receiving an indication identifying one of the hierarchically arranged semantic areas; and

obtaining, based on the attention mask, an attention score for the identified one of the hierarchically arranged semantic areas,

wherein updating the image comprises updating, based on the attention score for the identified one of the hierarchically arranged semantic areas, the identified one of the hierarchically arranged semantic areas of the image to visually identify an attention of the artificial intelligence model to the identified one of the hierarchically arranged semantic areas.

19. The method of claim 10 , further comprising:

receiving an indication identifying one of the hierarchically arranged semantic areas; and

modifying the attention mask to change the attention score for the identified one of the hierarchically arranged areas.

20. A non-transitory computer-readable storage medium having stored thereon instructions that, when executed, cause one or more processors to:

receive a query regarding an image;

segment the image into hierarchically arranged semantic areas in which one or more objects in the image are segmented into one or more parts;

determine, based on the query, an attention mask for the hierarchically arranged semantic areas in which an attention score is assigned to one or more of each of the one or more objects and each of the one or more parts;

update, based on the attention mask, the image to visually identify which of the hierarchically arranged semantic areas formed a basis for the result output by the artificial intelligence model; and

output the updated image.