Explaining Deep Classification of Time-Series Data with Learned Prototypes.

Area of Science:

• Artificial Intelligence
• Machine Learning
• Biomedical Signal Processing

Background:

• Deep learning models are increasingly used for high-risk decisions, but their "black box" nature hinders trust.
• Explainable AI (XAI) is crucial for understanding algorithmic decision-making processes.
• Latent space representations in deep learning offer potential for interpretable insights.

Purpose of the Study:

• To develop and evaluate an explainable AI framework using learned prototypes from deep learning models.
• To investigate the impact of prototype diversity and robustness on classification accuracy and interpretability.
• To demonstrate the ability of prototypes to capture meaningful real-world features in time-series data.

Main Methods:

• Leveraging latent space data to learn stereotypical representations (prototypes) during deep learning model training.
• Applying the prototype framework to two-dimensional time-series data, including electrocardiogram (ECG), respiration, and audio waveforms.
• Optimizing models for increased prototype diversity and robustness.
• Visualizing prototype usage in the latent space for class distinction.

Main Results:

• Prototypes effectively learned real-world features, such as bradycardia in ECG, apnea in respiration, and articulation in speech.
• The framework provided explainable insights into classification tasks across different domains.
• Optimized prototypes enhanced model performance and interpretability.
• Visualizations demonstrated how prototypes contribute to distinguishing between classes.

Conclusions:

• Learned prototypes offer a powerful method for achieving explainable AI in deep learning models.
• The prototypical framework is effective for analyzing two-dimensional time-series data, yielding clinically and technically relevant insights.
• This approach enhances user confidence in applying AI to high-stakes decision-making scenarios.