Systems and methods for machine learning enhanced by human measurements
In various embodiments, training objects are classified by human annotators, psychometric data characterizing the annotation of the training objects is acquired, a human-weighted loss function based at least in part on the classification data and the psychometric data is computationally derived, and one or more features of a query object are computationally classified based at least in part on the human-weighted loss function.
1. A computer-implemented method for data classification and identification, the method comprising:
providing, over a computer network, data corresponding to a plurality of training objects to a plurality of training devices each associated with one of a plurality of human annotators, each of the training objects comprising features for classification;
displaying the training objects on a display of each of the training devices;
receiving, via communication interfaces of at least some of the training devices, classification data comprising at least some of the training objects annotated, via annotation interfaces of the training devices, by at least some of the annotators with classifications for features thereof;
acquiring psychometric data characterizing the annotation of the training objects by the annotators;
computationally deriving a human-weighted loss function based at least in part on the classification data and the psychometric data, the loss function comprising penalties for misclassification, magnitudes of the penalties increasing with increasing deviation from the classification data;
receiving, by a classification device, data corresponding to a query object different from the plurality of training objects; and
thereafter, computationally classifying, by a computer processor, at least one feature of the query object based at least in part on the human-weighted loss function.
2. The method of claim 1 , further comprising displaying the classification of the at least one feature of the query object.
3. The method of claim 1 , wherein the magnitudes of the penalties for misclassification are based at least in part on the psychometric data.
4. The method of claim 1 , wherein computationally deriving the human-weighted loss function comprises:
computationally classifying, by the computer processor, at least some of the training objects based at least in part on an initial loss function, thereby generating training data;
comparing the training data to the classification data to identify, within the training data, features misclassified in comparison to the classification data;
assigning the penalties for misclassification to the misclassified features in the training data; and
incorporating the penalties for misclassification within the initial loss function to generate the human-weighted loss function.
5. The method of claim 4 , wherein the penalties for misclassification are assigned based at least in part on the psychometric data.
6. The method of claim 4 , wherein the initial loss function comprises a hinge loss function.
7. The method of claim 1 , wherein the psychometric data comprises at least one of (i) response time for classifying one or more features, (ii) accuracy of feature classification, or (iii) presentation time of one or more training objects.
8. The method of claim 1 , wherein each of the training objects comprises a digital image, and one or more features for classification comprise human faces.
9. The method of claim 1 , wherein at least one of the training devices comprises a computer or mobile computing device.
10. The method of claim 1 , wherein the query object comprises a digital image, and at least one said feature of the query object comprises a human face.
11. A system for data classification and identification, the system comprising:
a database of training objects, the database comprising a storage medium populated with stored computer records specifying, for each of a plurality of training objects, (i) classification data comprising annotations received from a plurality of human annotators, and (ii) psychometric data characterizing the annotation of the training object by the plurality of human annotators;
a computer processor;
a classification device for receiving query objects different from the training objects in the database;
a penalization module, executable by the computer processor, for deriving a human-weighted loss function based at least in part on the classification data in the database and the psychometric data of at least some of the training objects in the database, the loss function comprising penalties for misclassification, magnitudes of the penalties increasing with increasing deviation from the classification data; and
a classification module, executable by the computer processor, for classifying features of query objects based at least in part on the human-weighted loss function.
12. The system of claim 11 , further comprising a display module, executable by the computer processor, for displaying training objects to the plurality of human annotators.
13. The system of claim 12 , further comprising a plurality of training devices, each associated with a human annotator, for displaying training objects, each training device comprising a communication interface for receiving training objects and transmitting classification data.
14. The system of claim 13 , wherein at least one of the training devices comprises a computer or mobile computing device.
15. The system of claim 11 , wherein the psychometric data comprises at least one of (i) response time for classifying one or more features, (ii) accuracy of feature classification, or (iii) presentation time of one or more training objects.
16. The system of claim 11 , wherein each of the training objects comprises a digital image, and one or more features for classification comprise human faces.
17. The system of claim 11 , wherein at least one query object comprises a digital image, and at least one feature of the query object comprises a human face.