Device and method for generating a counterfactual data sample for a neural network

Abstract

Disclosed are a device and method for generating a counterfactual data sample for a neural network. The method for generating a counterfactual data sample for a neural network based on an input sensordata sample is described. The method includes determining, using the neural network, a class prediction for the input sensor data sample, determining, in addition to the class prediction, an estimateof the uncertainty of the class prediction, generating a candidate counterfactual data sample for which the neural network determines a different class prediction than for the input sensor data sample, determining a loss function, wherein the loss function includes the estimate of the uncertainty of the class prediction by the neural network for the candidate counterfactual data sample, modifyingthe candidate counterfactual data sample to obtain a counterfactual data sample based on the determined loss function and outputting the counterfactual data sample.

Description

technical field [0001] The present disclosure relates to methods and apparatus for generating counterfactual data samples for neural networks. Background technique [0002] Deep learning models using neural networks are becoming more widely used, however, understanding these models before they are deployed in the field, especially when they are applied to high-stakes tasks such as autonomous driving or medical diagnosis How their results (predictions) are derived is critical. [0003] To understand a model, it is important to be able to establish quantitatively how well it has learned the desired input-output relationship. However, deep learning models and techniques typically lack the metrics and practice to measure this effect, and often produce models that are overparameterized compared to the amount of data available. This is especially true for models used in classification tasks, where the large number of model parameters allows the decision boundaries between object...

AI Technical Summary

Problems solved by technology

These small changes, such as just a few pixels of the input image, may not be relevant to the part representing the semantic object, and thus may not be helpful in interpreting the classification score

Images