SENSITIVITY BASED MODEL AGNOSTIC SCALABLE EXPLANATIONS OF DEEP LEARNING.

Area of Science:

• Artificial Intelligence
• Machine Learning
• Bioinformatics

Background:

• Deep neural networks (DNNs) excel at prediction but lack transparency.
• Understanding DNNs' learned mechanisms is crucial for scientific validation and health applications.

Purpose of the Study:

• To develop SensX, a model-agnostic explainable AI (XAI) framework.
• To enhance the interpretability of DNNs in biological and clinical contexts.
• To improve upon existing XAI methods in accuracy, speed, and consistency.

Main Methods:

• Designed SensX as a model-agnostic XAI framework.
• Evaluated SensX against state-of-the-art XAI methods.
• Applied SensX to explain Vision Transformer (ViT) models and DNNs for single-cell RNA-seq data analysis.

Main Results:

• SensX achieved higher accuracy (up to 52%) and faster computation (up to 158x) than current XAI.
• Identified optimal subsets of input features, reducing dimensionality.
• Successfully explained large-scale ViT models and identified key genes for cell type annotation.

Conclusions:

• SensX provides a scalable and efficient solution for DNN interpretability.
• The framework validates learned features and reveals architectural biases.
• SensX facilitates hypothesis generation and model validation in data-driven science.