Accelerated training of a machine learning model
Systems and methods are presented for training a second machine learning model according to aspects of a trained first machine learning model. Processing features utilized by a training framework to train the first machine learning model are identified, and at least some of the processing features are combined with an initial set of training features to form updated training features. The updated training features are presented to a user for customization, resulting in customized training features. An executable training framework is configured with the customized training features and executed to train the second machine learning model.
1 . A computer-implemented method, comprising:
accessing a plurality of processing features of a first machine learning model previously trained to process input data of a corpus of input data, wherein the plurality of processing features were utilized by a training framework in training the first machine learning model;
determining a plurality of initial training features according to one or more analyses of input data of the corpus of input data for training a second machine learning model;
combining at least a portion of the plurality of processing features with at least a portion of the plurality of initial training features to form updated training features for training the second machine learning model, the combining comprising:
determining discrete processing features of the plurality of processing features that correspond to discrete training features of the plurality of initial training features; and
for each discrete processing feature having a corresponding discrete training feature:
combining values of the discrete training feature with values of the discrete processing feature to form an updated training feature;
customizing at least some of the updated training features to form customized training features;
incorporating the customized training features into an executable training framework for training the second machine learning model;
initializing the customized training features, wherein initializing includes warm-starting at least one feature of the customized training features from a processing feature of the first machine learning model; and
executing the executable training framework to train the second machine learning model according to at least some input data of the corpus of input data.
2 . The computer-implemented method of claim 1 , comprising:
determining the plurality of processing features correspond to first information that was associated with input data of the corpus of input data at a prior time period of training the first machine learning model; and
determining the plurality of initial training features correspond to second information currently associated with input data of the corpus of input data, where the first information and the second information are different.
3 . The computer implemented method of claim 2 , comprising determining the plurality of processing features according to the first information and determining the plurality of initial training features according to the second information.
4 . The computer-implemented method of claim 1 , comprising
determining that at least one feature of the plurality of processing features includes a first discrete feature associated with a first name and a first set of vocabulary terms; and
determining that at least one feature of the plurality of initial training features includes a second discrete feature associated with the first name and a second set of vocabulary terms; and
wherein the first set of vocabulary terms and the second set of vocabulary terms are different.
5 . The computer-implemented method of claim 1 , comprising:
determining that the plurality of processing features include one or more first embeddings associated with items of input data of the corpus of input data; and
determining that the plurality of initial training features includes a second embedding not included in the one or more first embeddings.
6 . The computer implemented method of claim 1 , wherein customizing at least some of the updated training features comprises:
receiving a user input identifying at least one feature of the updated training features; and
excluding the at least one feature of the updated training features to form the customized training features.
7 . The computer implemented method of claim 1 , wherein customizing at least some of the updated training features comprises:
receiving a user input identifying at least one feature not included in the updated training features; and
adding the at least one feature to the updated training features to form the customized training features.
8 . The computer-implemented method of claim 1 , wherein customizing at least some of the updated training features comprises:
receiving a user input modifying at least one feature of the updated training features; and
using the modified at least one feature to form the customized training features.
9 . A computer readable medium bearing computer executable instructions which, when executed by an online service operating on a computer system comprising at least a processor, carry out a method comprising:
accessing a plurality of processing features of a first machine learning model;
determining a plurality of initial training features for training a second machine learning model to process input data of a corpus of input data, wherein the plurality of initial training features are determined based at least in part on one or more analyses of input data of the corpus of input data;
combining at least a portion of the plurality of processing features with at least a portion of the plurality of initial training features to form updated training features for training the second machine learning model, the combining comprising:
determining discrete processing features of the plurality of processing features that correspond to discrete training features of the plurality of initial training features; and
for each discrete processing feature having a corresponding discrete training feature:
combining values of the discrete training feature with values of the discrete processing feature to form an updated training feature;
customizing at least some of the updated training features to form customized training features;
initializing the customized training features; and
training the second machine learning model utilizing the customized training features.
10 . The computer readable medium of claim 9 , further comprising instructions that when executed perform the method further comprising:
determining the plurality of processing features correspond to first information that was associated with input data of the corpus of input data at a prior time period of training the first machine learning model; and
determining the plurality of initial training features correspond to second information currently associated with input data of the corpus of input data, where the first information and the second information are different.
11 . The computer readable medium of claim 9 , further comprising instructions that when executed perform the method further comprising:
determining that at least one feature of the plurality of processing features includes a first discrete feature associated with a first name and a first set of vocabulary terms; and
determining that at least one feature of the plurality of initial training features includes a second discrete feature associated with the first name and a second set of vocabulary terms; and
wherein the first set of vocabulary terms and the second set of vocabulary terms are different.
12 . The computer readable medium of claim 9 , further comprising instructions that when executed perform the method further comprising:
determining that the plurality of processing features include one or more first embeddings associated with one or more items of input data of the corpus of input data; and
determining that the plurality of initial training features includes a second embedding not included one or more first embeddings.
13 . The computer readable medium of claim 9 , wherein the initializing comprises warm-starting at least one a feature of the customized training features from a processing feature of the first machine learning model.
14 . The computer readable medium of claim 9 , wherein customizing comprises at least one of:
receiving a user input excluding at least one feature of the updated training features to form the customized training features;
receiving a user input including at least one feature not included in the updated training features to form the customized training features; or
receiving a user input modifying at least one feature of the updated training features to form the customized training features.
15 . The computer-readable medium of claim 14 , wherein modifying at least one feature comprises: changing at least one parameter of the at least one feature of the updated training features.
16 . The computer readable medium of claim 9 , the method further comprising:
providing at least a portion of the updated training features to a user for customization;
receiving a customization indication from the user of the updated training features; and
wherein customizing is in response to receiving the customization indication and based at least in part on the customization indication.
17 . A computer system, comprising:
one or more processors; and
a memory storing program instructions that when executed by the one or more processors cause the one or more processors to at least:
access a plurality of processing features of a first machine learning model, wherein the first plurality of processing features were utilized by a training framework in training the first machine learning model;
determine a plurality of initial training features for training a second machine learning model to process input data of a corpus of input data, wherein the plurality of initial training features are determined according to one or more analyses of input data of the corpus of input data;
combine at least one feature of the plurality of processing features with at least one feature of the plurality of initial training features to form updated training features for training the second machine learning model, the combining comprising:
determining discrete processing features of the plurality of processing features that correspond to discrete training features of the plurality of initial training features; and
for each discrete processing feature having a corresponding discrete training feature:
combining values of the discrete training feature with values of the discrete processing feature to form an updated training feature;
customize the updated training features to form customized training features, wherein customization of the updated training features includes at least one of:
exclude at least one feature of the updated training features to form the customized training features;
add at least feature that was not included in the updated training features to form the customized training features; or
modify at least one feature of the updated training features to form the customized training features;
incorporate the customized training features into an executable training framework for training the second machine learning model;
initialize the customized training features, wherein initialization includes warm-starting at least one feature of the customized training features from a processing feature of the first machine learning model; and
execute the executable training framework to train the second machine learning model according to at least some input data of the corpus of input data.
18 . The computer system of claim 17 , wherein the program instructions that when executed by the one or more processors to modify the at least one feature of the updated training features further include instructions that when executed by the one or more processors further cause the one or more processors to at least:
alter at least one parameter of the at least one feature of the updated training features.
19 . The computer system of claim 17 , wherein the program instructions that when executed by the one or more processors further cause the one or more processors to at least:
receive a customization indication; and
wherein the customization is based at least in part on the customization indication.
20 . The computer system of claim 19 , further comprising:
receiving the customization indication from a user.