Towards Relatable Explainable AI with the Perceptual Process

Document Title

Towards Relatable Explainable AI with the Perceptual Process

Document Information

• Subject Area: Explainable Artificial Intelligence (XAI), particularly contrastive explanations in perceptual tasks
• Keywords: Explainable AI (XAI), contrastive explanations, speech emotion recognition, deep learning models, user needs, human cognition

Research Background and Issues

• Problems and Challenges:

  • Machine learning models are often highly complex, making them difficult for end users to understand, which limits their real-world applications.
  • Current contrastive explanation methods remain overly basic, lack semantic meaning, and are difficult to align with human cognition.
  • In the audio data domain, existing explanation methods (e.g., saliency maps based on spectrograms) are overly technical and challenging for general users to comprehend.

• Significance:

  • Explainability is a key factor in achieving responsible and trustworthy artificial intelligence.
  • Applications in the audio domain (e.g., speech emotion recognition) require explanation methods that align with human perceptual processes to enhance usability and trust.

• Research Motivation:

  • Inspired by the theory of human perceptual processes, this study proposes a more relatable explanation framework that aligns explanations with users' conceptual, hypothetical, and associative cognitive needs.
  • It explores various contrastive explanations, including saliency, counterfactuals, and supplementary cues.

Solution

• Proposed Framework and Model:

  • XAI Perceptual Processing Framework: Based on the theory of human perceptual processes, this framework is divided into three stages: Select, Organize, and Interpret.
  • RexNet Model (Relatable Explanation Network): A modular multi-task deep learning framework supporting three types of contrastive explanations:
    1. Contrastive Saliency: Generates contrastive saliency maps to help users understand the input features the model focuses on.
    2. Counterfactual Synthetic: Uses a generative adversarial network (StarGAN-VC) to generate style-transformed speech samples, showing differences from the target classification.
    3. Contrastive Cues: Provides emotion-related speech cues (e.g., high-frequency energy, volume) for comparison.

• Innovations:

  • Incorporates human cognitive psychology by introducing perceptual theories into AI explanation design.
  • Proposes a model architecture that organically integrates saliency, counterfactual examples, and cue information.
  • First application of relatable explanation methods in audio prediction tasks.

• Implementation Steps:

  • Train a speech emotion classification model using the RAVDESS dataset and extend it into RexNet, achieving emotion prediction through a spectrogram-based CNN.
  • Integrate explanation modules and generate explanations using techniques such as Grad-CAM, generative adversarial networks (GAN), and layer-wise relevance propagation (LRP).
  • Evaluate the usability and effectiveness of the explanations through modeling studies and user experiments.

Research Outcomes

• Specific Outcomes:

  • Established and validated the XAI Perceptual Processing Framework and the RexNet framework.
  • Proposed three forms of explanations—contrastive saliency maps, counterfactual example generation, and contrastive cues—suitable for speech emotion recognition tasks.
  • Experiments demonstrated that counterfactual examples and contrastive cue explanations are effective in improving user decision quality and trust.

• Advantages Compared to Existing Solutions:

  • Compared to existing explanation methods, RexNet’s multi-faceted contrastive explanations significantly enhance users' understanding of model predictions.
  • Unlike traditional saliency map explanations, counterfactual examples and contrastive cues are more directly aligned with human perceptual processes.

• Experimental and Evaluation Results:

  • Experiments showed RexNet achieved an initial emotion classification accuracy of 79.5%, a significant improvement over traditional CNN models (75.7%).
  • User experiments revealed that explanations combining counterfactual examples and contrastive cues effectively improved decision accuracy and trust in the model, while saliency explanations alone had limited impact.

• Limitations and Future Directions:

  • Limitations:
    • Current counterfactual example generation suffers from accuracy issues.
    • The interpretability of saliency information needs further refinement.
    • Using instance data (rather than generated speech) limits the diversity of counterfactual examples.
  • Future Research Directions:
    • Employ more advanced speech synthesis techniques (e.g., Seq2Seq-VC or StarGAN-VC v2) to improve the quality of counterfactual generation.
    • Explore ways to further refine the semantic content of saliency maps.
    • Apply this framework to other complex domains (e.g., medical diagnosis or industrial inspection) for perceptual task explanations.