Prompt-based attribution of generated media contents
Systems, methods and non-transitory computer readable media for identifying prompts used for training of inference models are provided. In some examples, a specific textual prompt in a natural language may be received. Further, data based on at least one parameter of an inference model may be accessed. The inference model may be a result of training a machine learning model using a plurality of training examples. Each training example of the plurality of training examples may include a respective textual content and a respective media content. The data and the specific textual prompt may be analyzed to determine a likelihood that the specific textual prompt is included in at least one training example of the plurality of training examples. A digital signal indicative of the likelihood that the specific textual prompt is included in at least one training example of the plurality of training examples may be generated.
1. A non-transitory computer readable medium storing a software program comprising data and computer implementable instructions that when executed by at least one processor cause the at least one processor to perform operations for prompt-based attribution of generated media contents to training examples, the operations comprising:
receiving a first media content generated using a generative model in response to a first textual input in a natural language, the generative model is a result of training a machine learning model using a plurality of training examples, each training example of the plurality of training examples includes a respective textual content in the natural language and a respective media content;
determining one or more properties of the first textual input;
using the one or more properties of the first textual input to attribute the first media content to a first subgroup of at least one but not all of the plurality of training examples;
determining that the training examples of the first subgroup are associated with a first at least one source; and
for each source of the first at least one source, updating a respective data-record associated with the source based on the attribution.
2. The non-transitory computer readable medium of claim 1 , wherein the operations further comprise:
determining one or more properties of the first media content; and
basing the attribution of the first media content to the first subgroup of at least one but not all of the plurality of training examples on the one or more properties of the first media content and the one or more properties of the first textual input.
3. The non-transitory computer readable medium of claim 1 , wherein the operations further comprise:
using the one or more properties of the first textual input to determine, for each training example of the first subgroup, a degree of attribution of the first media content to the respective training example; and
for each source of the first at least one source, further basing the update to the data-record associated with the source on at least one of the determined degrees.
4. The non-transitory computer readable medium of claim 1 , wherein the determination of the one or more properties of the first textual input is based on an intermediate result of the generative model when generating the first media content.
5. The non-transitory computer readable medium of claim 1 , wherein the training of the machine learning model to obtain the generative model includes an iterative process for reducing a loss function, wherein in each iteration of the iterative process a respective training example of the plurality of training examples is analyzed and the loss function is updated, and the determination of whether to attribute the first media content to a particular training example is based on the loss function in the iteration of the iterative process that includes the analysis of the particular training example.
6. The non-transitory computer readable medium of claim 1 , wherein the training of the machine learning model to obtain the generative model includes a first training step and a second training step, the first training step uses a second subgroup of the plurality of training examples to obtain an intermediate model, the second training step uses a third subgroup of the plurality of training examples and uses the intermediate model for initialization to obtain the generative model, the third subgroup differs from the second subgroup, and wherein the operations further comprise:
comparing a result associated with the first textual input and the intermediate model with a result associated with the first textual input and the generative model; and
for each training example of the third subgroup, determining whether to include the respective training example in the first subgroup based on a result of the comparison.
7. The non-transitory computer readable medium of claim 1 , wherein the operations further comprise:
receiving a second media content generated using the generative model in response to a second textual input;
determining one or more properties of the second textual input;
using the one or more properties of the second textual input to attribute the second media content to a second subgroup of at least one but not all of the plurality of training examples;
determining that the training examples of the second subgroup are associated with a second at least one source, the second at least one source includes one or more sources not included in the first at least one source;
based on the second at least one source, forgoing usage of the second media content; and
initiating usage of the first media content.
8. The non-transitory computer readable medium of claim 1 , wherein the operations further comprise:
receiving a second media content generated using the generative model in response to a second textual input;
determining one or more properties of the second textual input;
using the one or more properties of the second textual input to attribute the second media content to a second subgroup of at least one but not all of the plurality of training examples;
accessing a data-structure associating training examples with amounts;
using the data-structure to determine that the training examples of the first subgroup are associated with a first total amount;
using the data-structure to determine that the training examples of the second subgroup are associated with a second total amount;
based on the first and second total amounts, forgoing usage of the second media content; and
initiating usage of the first media content.
9. The non-transitory computer readable medium of claim 1 , wherein the operations further comprise:
accessing a data-structure associating training examples with amounts;
using the data-structure to determine that the training examples of the first subgroup are associated with a first total amount; and
further basing the updates to the data-records associated with the first at least one source on the first total amount.
10. The non-transitory computer readable medium of claim 1 , wherein the operations further comprise using a second machine learning model to analyze the one or more properties of the first textual input to determine whether to attribute the first media content to the particular training example.
11. The non-transitory computer readable medium of claim 1 , wherein the operations further comprise:
using the one or more properties of the first textual input to identify a mathematical object in a mathematical space; and
using the mathematical object to determine whether to attribute the first media content to a particular training example.
12. The non-transitory computer readable medium of claim 1 , wherein the first textual input includes a first noun, the textual content included in a particular training example includes a second noun, and wherein the operations further comprise using the first noun and the second noun to determine whether to attribute the first media content to the particular training example.
13. The non-transitory computer readable medium of claim 1 , wherein the first textual input includes a particular noun and a first adjective adjacent to the particular noun, the textual content included in a particular training example includes the particular noun and a second adjective adjacent to the particular noun, and wherein the operations further comprise using the first adjective and the second adjective to determine whether to attribute the first media content to the particular training example.
14. The non-transitory computer readable medium of claim 1 , wherein the first textual input includes a particular verb and a first adverb adjacent to the particular verb, the textual content included in a particular training example includes the particular verb and a second adverb adjacent to the particular verb, and wherein the operations further comprise using the first adverb and the second adverb to determine whether to attribute the first media content to the particular training example.
15. The non-transitory computer readable medium of claim 1 , wherein the first textual input includes a first word and a second word, the textual content included in a particular training example includes the first word and the second word, and wherein the operations further comprise determining whether to attribute the first media content to the particular training example based on an arrangement of the first word and the second word in the first textual input and based on an arrangement of the first word and the second word in the textual content included in the particular training example.
16. The non-transitory computer readable medium of claim 1 , wherein the one or more properties of the first textual input are indicative of a language register of the first textual input, and wherein the determination of whether to attribute the first media content to a particular training example is based on the language register of the first textual input.
17. The non-transitory computer readable medium of claim 1 , wherein the one or more properties of the first textual input are indicative of a subject matter, and wherein the determination of whether to attribute the first media content to the particular training example is based on the subject matter.
18. The non-transitory computer readable medium of claim 1 , wherein the one or more properties of the first textual input are indicative of a first source, and wherein the determination of whether to attribute the first media content to the particular training example is based on the first source.
19. A method for prompt-based attribution of generated media contents to training examples, the method comprising:
receiving a first media content generated using a generative model in response to a first textual input in a natural language, the generative model is a result of training a machine learning model using a plurality of training examples, each training example of the plurality of training examples includes a respective textual content in the natural language and a respective media content;
determining one or more properties of the first textual input;
using the one or more properties of the first textual input to attribute the first media content to a first subgroup of at least one but not all of the plurality of training examples;
determining that the training examples of the first subgroup are associated with a first at least one source; and
for each source of the first at least one source, updating a respective data-record associated with the source based on the attribution.
20. A system for prompt-based attribution of generated media contents to training examples, the system comprising:
at least one processor configured to perform the operations of: receiving a first media content generated using a generative model in response to a first textual input in a natural language, the generative model is a result of training a machine learning model using a plurality of training examples, each training example of the plurality of training examples includes a respective textual content in the natural language and a respective media content;
determining one or more properties of the first textual input;
using the one or more properties of the first textual input to attribute the first media content to a first subgroup of at least one but not all of the plurality of training examples;
determining that the training examples of the first subgroup are associated with a first at least one source; and
for each source of the first at least one source, updating a respective data-record associated with the source based on the attribution.