SENSITIVITY BASED MODEL AGNOSTIC SCALABLE EXPLANATIONS OF DEEP LEARNING

Deep neural networks (DNNs) are powerful tools for data-driven predictive machine learning, but their complex architecture obscures mechanistic relations that they have learned from data. This information is critical to the scientific method of hypotheses development, experiment design, and model validation, especially when DNNs are used for biological and clinical predictions that affect human health. We design SensX, a model agnostic explainable AI (XAI) framework that outperformed current state-of-the-art XAI in accuracy (up to 52% higher) and computation time (up to 158 times faster), with higher consistency in all cases. It also determines an optimal subset of important input features, reducing dimensionality of further analyses. SensX scaled to explain vision transformer (ViT) models with more than 150,000 features, which is computationally infeasible for current state-of-the-art XAI. SensX validated that ViT models learned justifiable features as important for different facial attributes of different human faces. SensX revealed biases inherent to the ViT architecture, an observation possible only when importance of each feature is explained. We trained DNNs to annotate biological cell types using single-cell RNA-seq data and SensX determined the sets of genes that the DNNs learned to be important to different cell types.