System and method for anomaly detection via a multi-prediction-model architecture

Abstract

In some embodiments, anomaly detection may be facilitated via a multi-neural-network architecture. In some embodiments, a first neural network may be configured to generate hidden representations of data items corresponding to a concept. A second neural network may be configured to generate reconstructions of the data items from the hidden representations. The first neural network may be configured to assess the reconstructions against the data items and update configurations of the first neural network based on the assessment of the reconstructions. Subsequent to the update of the first neural network, the first neural network may generate a hidden representation of a first data item from the first data item. The second neural network may generate a reconstruction of the first data item from the hidden representation. An anomaly in the first data item may be detected based on differences between the first data item and the reconstruction.

1 - 20 . (canceled) 
     
     
         21 . A method of facilitating anomaly detection via a multi-model architecture, the method being implemented by one or more processors executing computer program instructions that, when executed, perform the method, the method comprising:
 obtaining data items that corresponds to a concept;   providing the data items to a first model to cause the first model to generate hidden representations of the data items from the data items;   providing the hidden representations of the data items to a second model to cause the second model to generate reconstructions of the data items from the hidden representations of the data items;   providing the reconstructions of the data items to the first model, the first model updating one or more representation-generation-related configurations of the first model based on the data items and the reconstructions of the data items.   
     
     
         22 . The method of  claim 21 , further comprising: subsequent to providing the reconstructions of the data items, performing the following operations:
 providing a given data item to the first model to cause the first model to generate a hidden representation of the given data item from the given data item; and   providing the hidden representation of the given data item to the second model to cause the second model to generate a reconstruction of the given data item from the hidden representation of the given data item,   wherein no anomaly is detected in the given data item based on differences between the given data item and the reconstruction of the given data item.   
     
     
         23 . The method of  claim 21 , further comprising: subsequent to providing the reconstructions of the data items, performing the following operations:
 providing a given data item to the first model to cause the first model to generate a hidden representation of the given data item from the given data item;   providing the hidden representation of the given data item to the second model to cause the second model to generate a reconstruction of the given data item from the hidden representation of the given data item; and   detecting an anomaly in the given data item based on differences between the given data item and the reconstruction of the given data item.   
     
     
         24 . The method of  claim 23 , further comprising: subsequent to providing the reconstructions of the data items, performing the following operations:
 obtaining additional data items that corresponds to the concept;   providing the additional data items to the first model to cause the first model to generate hidden representations of the additional data items from the additional data items;   providing the hidden representations of the additional data items to the second model to cause the second model to generate reconstructions of the additional data items from the hidden representations of the additional data items;   providing the additional data items, the reconstructions of the additional data items, and given reference feedback to a third model to cause the third model to be trained based on the additional data items, the reconstructions of the additional data items, and the given reference feedback to generate an indication that each additional data item of the additional data items and the reconstruction corresponding to the additional data item are similar; and   providing the given data item and the reconstruction of the given data item to the third model to cause the third model to assess the differences between the given data item and the reconstruction of the given data item, the third model generating an indication that the given data item and the reconstruction of the given data item are not similar based on the differences between the given data item and the reconstruction of the given data item,   wherein detecting the anomaly comprises detecting the anomaly in the given data item based on the indication generated by the third model.   
     
     
         25 . The method of  claim 23 , wherein the first model is configured to generate additional hidden representations of the data items from the data items subsequent to the updating of the first model, the method further comprising:
 providing the additional hidden representations of the data items to the second model to cause the second model to generate additional reconstructions of the data items from the additional hidden representations of the data items; and   providing the additional reconstructions of the data items as reference feedback to the first model to cause the first model to assess the additional reconstructions of the data items against the data items, the first model further updating one or more representation-generation-related configurations of the first model based on the first model's assessment of the additional reconstructions of the data items.   
     
     
         26 . The method of  claim 25 , further comprising:
 providing the data items, the additional reconstructions of the data items, and given reference feedback to a third model to cause the third model to be trained based on the data items, the additional reconstructions of the data items, and the given reference feedback to generate an indication that each data item of the data items and the additional reconstruction corresponding to the data item are similar; and   providing the given data item and the reconstruction of the given data item to the third model to cause the third model to assess the differences between the given data item and the reconstruction of the given data item, the third model generating an indication that the given data item and the reconstruction of the given data item are not similar based on the differences between the given data item and the reconstruction of the given data item,   wherein detecting the anomaly comprises detecting the anomaly in the given data item based on the indication generated by the third model.   
     
     
         27 . The method of  claim 26 ,
 wherein the third model generates one or more indications of which portions of the given data item and the reconstruction of the given data item are not similar, and   wherein detecting the anomaly comprises detecting the anomaly in the given data item based on the one or more indications generated by the third model.   
     
     
         28 . The method of  claim 27 ,
 wherein the third model generates one or more additional indications of which portions of the given data item and the reconstruction of the given data item are similar, and   wherein detecting the anomaly comprises detecting the anomaly in the given data item based on the one or more indications and the one or more additional indications generated by the third model.   
     
     
         29 . The method of  claim 25 , further comprising:
 determining pairs such that each of the pairs comprises one of the data items and the additional reconstruction of another one of the data items;   providing the pairs to a third model to cause the third model to, with respect to each of the pairs, generate an indication of whether the corresponding data item and additional reconstruction of the pair are similar;   providing given reference feedback to the third model to cause the third model to assess the generated indications against the given reference feedback, the given reference feedback indicating that the corresponding data item and additional reconstruction of each of the pairs are not similar, the third model updating one or more configurations of the third model based on the third model's assessment of the generated indications; and   providing the given data item and the reconstruction of the given data item to the third model to cause the third model to assess the differences between the given data item and the reconstruction of the given data item, the third model generating an indication that the given data item and the reconstruction of the given data item are not similar based on the differences between the given data item and the reconstruction of the given data item,   wherein detecting the anomaly comprises detecting the anomaly in the given data item based on the indication generated by the third model.   
     
     
         30 . The method of  claim 21 , further comprising:
 determining subsets of data items such that each of the data item subsets comprise at least two data items of the data items;   providing the data item subsets to a third model to cause the third model to, with respect to each of the data item subsets, generate an indication of whether the two data items of the data item subset are similar;   providing given reference feedback to the third model to cause the third model to assess the generated indications against the given reference feedback, the given reference feedback indicating that the two data items of each of the data item subsets are not similar, the third model updating one or more configurations of the third model based on the third model's assessment of the generated indications; and   providing the given data item and the reconstruction of the given data item to the third model to cause the third model to assess the differences between the given data item and the reconstruction of the given data item, the third model generating an indication that the given data item and the reconstruction of the given data item are not similar based on the differences between the given data item and the reconstruction of the given data item,   wherein detecting the anomaly comprises detecting the anomaly in the given data item based on the indication generated by the third model.   
     
     
         31 . The method of  claim 21 , further comprising:
 deemphasizing one or more of the differences between the given data item and the reconstruction of the given data item,   wherein detecting the anomaly comprises detecting the anomaly in the given data item based on the one or more deemphasized differences and one or more other ones of the differences between the given data item and the reconstruction of the given data item.   
     
     
         32 . The method of  claim 21 , further comprising:
 emphasizing one or more of the differences between the given data item and the reconstruction of the given data item,   wherein detecting the anomaly comprises detecting the anomaly in the given data item based on the one or more emphasized differences and one or more other ones of the differences between the given data item and the reconstruction of the given data item.   
     
     
         33 . The method of  claim 21 , further comprising:
 deemphasizing one or more of the differences between the given data item and the reconstruction of the given data item; and   emphasizing one or more other ones of the differences between the given data item and the reconstruction of the given data item,   wherein detecting the anomaly comprises detecting the anomaly in the given data item based on the one or more deemphasized differences and the one or more emphasized differences.   
     
     
         34 . A system comprising:
 one or more processors executing computer program instructions that, when executed, cause the one or more processors to:
 obtain data items that corresponds to a concept; 
 provide the data items to a first model to cause the first model to generate hidden representations of the data items from the data items; 
 provide the hidden representations of the data items to a second model to cause the second model to generate reconstructions of the data items from the hidden representations of the data items; and 
 provide the reconstructions of the data items to the first model, the first model updating one or more representation-generation-related configurations of the first model based on the reconstructions of the data items. 
   
     
     
         35 . The system of  claim 34 , wherein the first model is configured to generate additional hidden representations of the data items from the data items subsequent to the updating of the first model, wherein the one or more processors are caused to:
 provide the additional hidden representations of the data items to the second model to cause the second model to generate additional reconstructions of the data items from the additional hidden representations of the data items; and   provide the additional reconstructions of the data items as reference feedback to the first model to cause the first model to assess the additional reconstructions of the data items against the data items, the first model further updating one or more representation-generation-related configurations of the first model based on the first model's assessment of the additional reconstructions of the data items.   
     
     
         36 . The system of  claim 35 , wherein the one or more processors are caused to:
 provide the data items, the additional reconstructions of the data items, and given reference feedback to a third model to cause the third model to be trained based on the data items, the additional reconstructions of the data items, and the given reference feedback to generate an indication that each data item of the data items and the additional reconstruction corresponding to the data item are similar; and   provide a given data item to the first model to cause the first model to generate a hidden representation of the given data item from the given data item;   provide the hidden representation of the given data item to the second model to cause the second model to generate a reconstruction of the given data item from the hidden representation of the given data item;   provide the given data item and the reconstruction of the given data item to the third model to cause the third model to assess the differences between the given data item and the reconstruction of the given data item, the third model generating an indication that the given data item and the reconstruction of the given data item are not similar based on the differences between the given data item and the reconstruction of the given data item; and   detecting an anomaly in the given data item based on the indication generated by the third model.   
     
     
         37 . A system comprising:
 a first model configured to generate hidden representations of data items from the data items, the data items corresponding to a concept;   a second model configured to generate reconstructions of the data items from the hidden representations of the data items;   wherein the first model is configured to:
 obtain the reconstructions of the data items; and 
 update one or more representation-generation-related configurations of the first model based on the reconstructions of the data items. 
   
     
     
         38 . The system of  claim 37 ,
 wherein, subsequent the update of the first model, the first model is configured to:   generate additional hidden representations of the data items from the data items;   wherein the second model is configured to generate additional reconstructions of the data items from the additional hidden representations of the data items; and   wherein the first model is configured to:
 assess the additional reconstructions of the data items against the data items; and 
 further update one or more representation-generation-related configurations of the first model based on the assessment of the additional reconstructions of the data items; and 
 subsequent the further update of the first model, generate a hidden representation of a given data item from the given data item; and 
   wherein the second model is configured to generate a reconstruction of the given data item from the hidden representation of the given data item; and   wherein the system comprises at least one processor configured to detect an anomaly in the given data item based on differences between the given data item and the reconstruction of the given data item.   
     
     
         39 . The system of  claim 38 , further comprising a third model configured to:
 update one or more configurations of the third model based on (i) the data items, (ii) the additional reconstructions of the data items, and (iii) reference feedback indicating that each data item of the data items and the additional reconstruction corresponding to the data item are similar; and   generate an indication that the given data item and the reconstruction of the given data item are not similar based on differences between the given data item and the reconstruction of the given data item; and   wherein detecting the anomaly comprises detecting the anomaly in the given data item based on the indication generated by the third model.   
     
     
         40 . The system of  claim 39 ,
 wherein the third model is configured to generate one or more indications of which portions of the given data item and the reconstruction of the given data item are not similar, and   wherein detecting the anomaly comprises detecting the anomaly in the given data item based on the one or more indications generated by the third model.