Joint input pertubation and temperature scaling for neural network calibration

A computer-readable medium storing instructions that, when executed by a computer system, cause the computer system to:

• retrieve a calibration training data set for a neural network that includes a plurality of data samples stored in an electronic file format;
• determine, based on the calibration training data set, a universal perturbation value associated with the plurality of data samples;
• determine a neural network classifier for a plurality of classes based on the plurality of data samples;
• determine a plurality of logits based on the neural network classifier, the plurality of logits including a respective logit associated with each respective class from the plurality of classes;
• determine, based on the plurality of logits and the universal perturbation value, a temperature scaling parameter associated with the plurality of classes;
• determine a respective probability score for each respective class in the plurality of classes based on the plurality of logits;
• and
• process, using the neural network, a testing data set that includes the plurality of data samples stored in the electronic file format to generate a respective calibrated probability score for each respective class in the plurality of classes, wherein processing the testing data set includes applying the universal perturbation value to the testing data set, and applying the temperature scaling parameter to the plurality of logits determined by the neural network based on the testing data set.

A computer-implemented method comprising:

• retrieving, by a computer system, a calibration training data set for a neural network that includes a plurality of data samples stored in an electronic file format;
• determining, by the computer system, based on the calibration training data set, a universal perturbation value associated with the plurality of data samples;
• determining, by the computer system, a neural network classifier for a plurality of classes based on the plurality of data samples;
• determining, by the computer system, a plurality of logits based on the neural network classifier, the plurality of logits including a respective logit associated with each respective class from the plurality of classes;
• determining, by the computer system, based on the plurality of logits and the universal perturbation value, a temperature scaling parameter associated with the plurality of classes;
• determining, by the computer system, a respective probability score for each respective class in the plurality of classes based on the plurality of logits;
• and
• processing, by the computer system using the neural network, a testing data set that includes the plurality of data samples stored in the electronic file format to generate a respective calibrated probability score for each respective class in the plurality of classes, wherein processing the testing data set includes applying the universal perturbation value to the testing data set, and applying the temperature scaling parameter to the plurality of logits determined by the neural network based on the testing data set.