מושגי ליבה
Developing DeepCSHAP for complex-valued neural networks to provide explainability.
תקציר
The paper introduces DeepCSHAP, adapting explanation methods for complex-valued neural networks. It addresses the lack of interpretability in real and complex-valued deep learning architectures. The authors develop a complex-valued variant of the DeepSHAP algorithm and adapt gradient-based explanation methods to the complex domain. They evaluate these methods on MNIST and PolSAR datasets, showing superior performance of DeepCSHAP. Theoretical results are validated, demonstrating fulfillment of SHAP properties.
סטטיסטיקה
"The model obtains a 88% accuracy on the digit classification task on the test set of MNIST."
"The trained model obtains an accuracy of 93% on the multiclass classification task using PolSAR dataset."