Synthetic Toy Case
Below you will find a link to exemplary samples of each synthetic data class with additional information about the generation process, and a link to the disentangled explanations extracted with DRSA for each class that reveal the decision behaviour of the network.
About the Synthetic Dataset
The synthetic dataset was specifically designed for quantitative evaluation of disentangled explanations. By creating audio samples with controlled properties, we can establish ground truth for concept-based explanations and rigorously evaluate the performance of DRSA.
This controlled environment allows for systematic analysis of how well the neural network identifies and separates distinct audio concepts, providing insights into the model's decision-making process that complement the real-world music genre recognition results.