MANOLO xSTAE: A generative framework

Demokritos “NCSR”worked closely with FDI to develop xSTAE, a generative framework for explaining classifier decisions through class-conditional autoencoders. The framework creates counterfactual explanations, by restyling misclassified inputs into versions that the classifier would assign to the correct class. This is achieved by training one autoencoder per class, each learning to reconstruct inputs in a way that shifts them toward the target class. By comparing the original input with its counterfactual version, one can identify which features the classifier expected to be more prominent, in order to make a correct prediction. In this collaboration, NCSR “D” developed the autoencoder-based XAI method, while FDI provided the classifier to be explained.

The framework was applied to EEG recordings from the Bitbrain EEG dataset (brainwave recordings collected during sleep using wearable headbands) and evaluated based on whether the restyled signals enhanced the key morphological features of each sleep stage, as defined by the experts, that the classifier had missed in the original signals.

The results demonstrate that the framework restyles EEG signals in a way that highlights the distinguishing characteristics the classifier associates with each class. Quantitatively, the restyle signals are much more likely to be correctly classified. Qualitatively, the comparisons between original and restyled signals reveal which features were underrepresented, offering insights into what the classifier needed to see more pronounced to change its decision.

Explaining classifiers’ decision-making is crucial for building trust and transparency, especially in high-stakes domains like healthcare. XAI helps experts understand, validate, and enhance system robustness, ensuring reliable performance even on data that is more difficult to classify.

The xSTAE framework supports T5.3: Explainability & Robustness Validation, part of the Trustworthy Efficiency & Performance Assessment component within MANOLO WP5. This task focuses on developing and validating methods to improve model explainability and robustness.

This outcome is significant for MANOLO because it:

• Offers visual explanations that clearly and intuitively illustrate AI systems’ decision-making.
• Provides a flexible framework for all MANOLO WPs to interpret and validate the robustness of their models.

This framework contributes to building reliable systems that meet the EU’s standards for transparency and accountability.