Exceeding the limits of visual-linguistic multi-task learning
Methods, computer readable media, and devices for exceeding the limits of visual-linguistic multi-task learning are disclosed. One method may include identifying a multi-modal multi-task classification dataset including a plurality of data examples, creating a transformer machine learning model to predict a plurality of categorical attributes of a product, and training the transformer machine learning model based on the multi-modal multi-task classification dataset using an alpha decay schedule and dynamically allocating task-specific parameters for at least one of the plurality of task-specific classification heads based on task complexity.
1. A computer-implemented method for exceeding the limits of visual-linguistic multi-task learning (MTL), the method comprising:
identifying a multi-modal multi-task unified ontology classification dataset comprising a plurality of data examples, at least one data example associated with a product and comprising textual data describing the associated product and a plurality of images of the associated product;
creating a MTL-based single-stream transformer machine learning model to predict a plurality of categorical attributes of a product, the MTL-based single-stream transformer machine learning model comprising:
a preprocessing layer forming a single input sequence;
a transformer layer; and
an output layer comprising a plurality of task-specific classification heads, a task representing a prediction of a categorical attribute; and
training the visual-linguistic MTL-based single-stream transformer machine learning model, simultaneously across all tasks, based on the multi-modal multi-task unified ontology classification dataset using an alpha decay schedule and dynamically allocating task-specific parameters for at least one of the plurality of task-specific classification heads based on task complexity, leveraging positive inductive transfer, from related tasks, for low-resource tasks.
2. The computer-implemented method of claim 1 , further comprising utilizing the trained transformer machine learning model to predict categorical attributes of a product.
3. The computer-implemented method of claim 1 , wherein dynamically allocating task-specific parameters for at least one of the plurality of task-specific classification heads based on task complexity comprises:
assigning a normally-distributed complexity score to at least one task;
dividing the plurality of task-specific classification heads into four quartiles based on the normally-distributed complexity score; and
assigning a first number of task-specific parameters to a first quartile of task-specific classification heads, a second number of task-specific parameters to a second quartile of task-specific classification heads, a third number of task-specific parameters to a third quartile of task-specific classification heads, and a fourth number of task-specific parameters to a fourth quartile of task-specific classification heads, wherein the first number is less than the second number, the second number is less than the third number, and the third number is less than the fourth number.
4. The computer-implemented method of claim 3 , wherein:
the first number of task-specific parameters is 64;
the second number of task-specific parameters is 128;
the third number of task-specific parameters is 512; and
the fourth number of task-specific parameters is 1024.
5. The computer-implemented method of claim 1 , wherein:
the plurality of data examples comprises data examples from a plurality of ecommerce websites; and
for at least one product offered by at least two of the plurality of ecommerce websites, the plurality of categorical attributes comprises at least a first categorical attribute having a different value for the at least two of the plurality of ecommerce websites and at least a second categorical attribute having a different value for the at least two of the plurality of ecommerce websites.
6. The computer-implemented method of claim 1 , wherein the plurality of categorical attributes includes one or more categorical attributes selected from the list comprising:
type; and
category.
7. The computer-implemented method of claim 1 , wherein the alpha decay schedule is selected from the list comprising:
demon;
linear;
cosine; and
exponential.
8. The computer-implemented method of claim 1 , wherein the transformer model comprises a bidirectional encoder representations from transformers architecture.
9. A non-transitory machine-readable storage medium that provides instructions that, if executed by a processor, are configurable to cause the processor to perform operations comprising:
identifying a multi-modal multi-task unified ontology classification dataset comprising a plurality of data examples, at least one data example associated with a product and comprising textual data describing the associated product and a plurality of images of the associated product;
creating a multi-task learning (MTL)-based single-stream transformer machine learning model to predict a plurality of categorical attributes of a product, the MTL-based single-stream transformer machine learning model comprising:
a preprocessing layer forming a single input sequence;
a transformer layer; and
an output layer comprising a plurality of task-specific classification heads, a task representing a prediction of a categorical attribute; and
training the visual-linguistic MTL-based single-stream transformer machine learning model, simultaneously across all tasks, based on the multi-modal multi-task unified ontology classification dataset using an alpha decay schedule and dynamically allocating task-specific parameters for at least one of the plurality of task-specific classification heads based on task complexity, leveraging positive inductive transfer, from related tasks, for low-resource tasks.
10. The non-transitory machine-readable storage medium of claim 9 , further comprising utilizing the trained transformer machine learning model to predict categorical attributes of a product.
11. The non-transitory machine-readable storage medium of claim 9 , wherein dynamically allocating task-specific parameters for at least one of the plurality of task-specific classification heads based on task complexity comprises:
assigning a normally-distributed complexity score to at least one task;
dividing the plurality of task-specific classification heads into four quartiles based on the normally-distributed complexity score; and
assigning a first number of task-specific parameters to a first quartile of task-specific classification heads, a second number of task-specific parameters to a second quartile of task-specific classification heads, a third number of task-specific parameters to a third quartile of task-specific classification heads, and a fourth number of task-specific parameters to a fourth quartile of task-specific classification heads, wherein the first number is less than the second number, the second number is less than the third number, and the third number is less than the fourth number.
12. The non-transitory machine-readable storage medium of claim 11 , wherein:
the first number of task-specific parameters is 64;
the second number of task-specific parameters is 128;
the third number of task-specific parameters is 512; and
the fourth number of task-specific parameters is 1024.
13. The non-transitory machine-readable storage medium of claim 9 , wherein:
the plurality of data examples comprises data examples from a plurality of ecommerce websites; and
for at least one product offered by at least two of the plurality of ecommerce websites, the plurality of categorical attributes comprises at least a first categorical attribute having a different value for the at least two of the plurality of ecommerce websites and at least a second categorical attribute having a different value for the at least two of the plurality of ecommerce websites.
14. The non-transitory machine-readable storage medium of claim 9 , wherein the transformer model comprises a bidirectional encoder representations from transformers architecture.
15. An apparatus comprising:
a processor; and
a non-transitory machine-readable storage medium that provides instructions that, if executed by a processor, are configurable to cause the processor to perform operations comprising:
identifying a multi-modal multi-task unified ontology classification dataset comprising a plurality of data examples, at least one data example associated with a product and comprising textual data describing the associated product and a plurality of images of the associated product;
creating a multi-task learning (MTL)-based single-stream transformer machine learning model to predict a plurality of categorical attributes of a product, the MTL-based single-stream transformer machine learning model comprising:
a preprocessing layer forming a single input sequence;
a transformer layer; and
an output layer comprising a plurality of task-specific classification heads, a task representing a prediction of a categorical attribute; and
training the visual-linguistic MTL-based single-stream transformer machine learning model, simultaneously across all tasks, based on the multi-modal multi-task unified ontology classification dataset using an alpha decay schedule and dynamically allocating task-specific parameters for at least one of the plurality of task-specific classification heads based on task complexity, leveraging positive inductive transfer, from related tasks, for low-resource tasks.
16. The apparatus of claim 15 , wherein dynamically allocating task-specific parameters for at least one of the plurality of task-specific classification heads based on task complexity comprises:
assigning a normally-distributed complexity score to at least one task;
dividing the plurality of task-specific classification heads into four quartiles based on the normally-distributed complexity score; and
assigning a first number of task-specific parameters to a first quartile of task-specific classification heads, a second number of task-specific parameters to a second quartile of task-specific classification heads, a third number of task-specific parameters to a third quartile of task-specific classification heads, and a fourth number of task-specific parameters to a fourth quartile of task-specific classification heads, wherein the first number is less than the second number, the second number is less than the third number, and the third number is less than the fourth number.
17. The apparatus of claim 16 , wherein:
the first number of task-specific parameters is 64;
the second number of task-specific parameters is 128;
the third number of task-specific parameters is 512; and
the fourth number of task-specific parameters is 1024.
18. The apparatus of claim 15 , wherein:
the plurality of data examples comprises data examples from a plurality of ecommerce websites; and
for at least one product offered by at least two of the plurality of ecommerce websites, the plurality of categorical attributes comprises at least a first categorical attribute having a different value for the at least two of the plurality of ecommerce websites and at least a second categorical attribute having a different value for the at least two of the plurality of ecommerce websites.
19. The apparatus of claim 15 , wherein the alpha decay schedule is selected from the list comprising:
demon;
linear;
cosine; and
exponential.
20. The apparatus of claim 15 , wherein the transformer model comprises a bidirectional encoder representations from transformers architecture.