Methods and apparatus for asynchronous and interactive machine learning using attention selection techniques

Abstract

A non-transitory medium includes code representing processor-executable instructions; the code causes a processor to produce, via a machine learning model, a predicted value of a membership relationship between a data object and a target tag. The code causes the processor to display, via a user interface, the data object and the target tag and indicate a non-empty set of identified sections of one or more attributes of data object supporting the membership relationship between the data object and the target tag. The code also causes the processor to receive a tag signal, via the user interface, indicating one of an acceptance tag signal, a dismissal tag signal, or a corrective tag signal, and re-train the machine learning model based at least in part on the tag signal.

1 . A non-transitory medium storing code representing a plurality of processor-executable instructions, the code comprising code to cause the processor to:
 produce, via a trained machine learning model, a predicted value for a membership relation between a data object and a target tag;   display, via a user interface, the data object and the target tag;   indicate a non-empty set of identified sections of one or more attribute values of the data object supporting the membership relation between the data object and the target tag;   receive a tag signal, via the user interface, indicating one of an acceptance tag signal, a dismissal tag signal, or a corrective tag signal; and   re-train the trained machine learning model based at least in part on the tag signal.   
     
     
         2 . The non-transitory medium of  claim 1 , wherein the non-empty set of identified sections of the one or more attribute values includes at least one of a set of text sections, a set of image sections, a set of video sections, or a set of metadata sections. 
     
     
         3 . The non-transitory medium of  claim 1 , wherein the target tag is generated by the trained machine learning model. 
     
     
         4 . The non-transitory medium of  claim 1 , wherein the code includes code to further cause the processor to:
 display, via the user interface, at least one salience value paired with an identified section from the non-empty set of identified sections of the one or more attributes of the data object.   
     
     
         5 . The non-transitory medium of  claim 1 , wherein the tag signal is a first tag signal, the target tag is a first target tag, and the code includes code to further cause the processor to:
 receive a second tag signal, via the user interface, indicating a user annotation of a section of an attribute value, the user annotation associated with a second target tag; and   re-train the trained machine learning model based at least in part on the second tag signal.   
     
     
         6 . The non-transitory medium of  claim 1 , wherein the code to produce via the trained machine learning model the predicted value includes code to:
 produce the predicted value as a function of at least one pre-salience value of at least one neuron of a neural network included in the machine learning model, the at least one neuron logically related with an attribute of the data object.   
     
     
         7 . The non-transitory medium of  claim 1 , wherein the code to indicate the non-empty set of identified sections of the one or more attribute values includes code to:
 determine a set of sections of the one or more attribute values of the data object supporting the membership relation, each section from the set of sections paired with a pre-salience value calculated as a function of a stochastic gradient descent between the data object and the target tag;   select from the set of sections of the one or more attribute values of the data object at least one section paired with a pre-salience value greater than a salience threshold value; and   send a signal to display, via the user interface, a graphical indicator highlighting the at least one section paired with a section salience value.   
     
     
         8 . The non-transitory medium of  claim 1 , wherein the code to indicate the non-empty set of identified sections of the one or more attribute values includes code to:
 identify a section of a value of a spatially decomposable attribute of the data object, the section confined to a finite spatial extent smaller than the spatial extent of the value of the spatially decomposable attribute; and   send a signal to display , via the user interface, a graphical indicator highlighting the identified section.   
     
     
         9 . The non-transitory medium of  claim 1 , wherein the code to indicate the non-empty set of identified sections of the one or more attribute values includes code to:
 identify a section of a value of a non-spatially decomposable attribute of the data object, the section confined to a finite spatial extent corresponding to the spatial extent of the value of the non-spatially decomposable attribute; and   display, via the user interface, a graphical indicator highlighting the identified section.   
     
     
         10 . The non-transitory medium of  claim 1 , wherein the machine learning model includes an attention neural network, the tag signal is an acceptance tag signal, and the code to re-train the trained machine learning model includes code to:
 reinforce positively at least one attention gate from the attention neural network, the at least one attention gate logically associated with a section from the non-empty set of identified sections of the one or more attribute values of the data object.   
     
     
         11 . The non-transitory medium of  claim 1 , wherein the machine learning model includes an attention neural network, the tag signal is a dismissal tag signal, and the code to re-train the trained machine learning model includes code to:
 reinforce negatively at least one attention gate from the attention neural network, the at least one attention gate logically associated with a section from the non-empty set of identified sections of the one or more attribute values of the data object.   
     
     
         12 . The non-transitory medium of  claim 1 , wherein the tag signal is an acceptance tag and the code to re-train the trained machine learning model includes code to:
 produce a pseudo-document upon receiving the tag signal, the pseudo-document including the non-empty set of identified sections indicating a positive membership relation between the data object and the target tag; and   re-train the trained machine learning model with a training set including the pseudo-document.   
     
     
         13 . The non-transitory medium of  claim 1 , wherein the tag signal is a dismissal tag signal and the code to re-train the machine learning model includes code to:
 produce a pseudo-document upon receiving the tag signal, the pseudo-document including the non-empty set of salient regions indicating a negative membership relation between the data object and the target tag; and   re-train the trained machine learning model with a training set including the pseudo-document.   
     
     
         14 . The non-transitory medium of  claim 1 , wherein the non-empty set of identified sections is a first non-empty set of identified sections, the tag signal is a corrective tag signal, and the code to re-train the machine learning model includes code to:
 produce a pseudo-document upon receiving the tag signal, the pseudo-document including a second non-empty set of identified sections that is different from the first non-empty set of identified sections; and   re-train the trained machine learning model with a training set including the pseudo-document.   
     
     
         15 . The non-transitory medium of  claim 1 , wherein the machine learning model includes an attention neural network, and the code to re-train the machine learning model includes code to:
 re-train the machine learning model based at least in part on the tag signal and a loss regularization process enforcing a user salience judgement received in the tag signal.   
     
     
         16 . A method comprising:
 producing, via a trained machine learning model, a predicted value for a membership relation between a data object and a target tag;   displaying, via a user interface, the data object and the target tag;   indicating a non-empty set of identified sections of one or more attribute values of the data object supporting the membership relation between the data object and the target tag;   receiving a tag signal, via the user interface, that indicates one of an acceptance tag signal, a dismissal tag signal, or a corrective tag signal; and   re-training the trained machine learning model based at least in part on the tag signal.   
     
     
         17 . The method of  claim 16 , wherein the predicted value indicates a probability that a user will annotate the data object with the target tag. 
     
     
         18 . The method of  claim 16 , wherein the tag signal indicates one of an acceptance, a dismissal, or a correction of a machine-generated judgement. 
     
     
         19 . The method of  claim 16 , further comprising:
 sending a signal, via the user interface, to display machine-generated judgements associated with each identified section of the one or more attribute values of the data object.   
     
     
         20 . The method of  claim 16 , wherein the trained machine learning model includes an attention neural network and re-training the trained machine learning model includes:
 updating at least one attention probabilistic gate included in an attention pooling layer of the attention neural network.