A Subspace Projection Based Technique for Visualizing Machine Learning Models

Ziqian Bi, Raymond Gao, Shiaofen Fang

Abstract: As Artificial Intelligence (AI) technology, particularly Machine Learning (ML) algorithms, becomes increasingly ubiquitous, our abilities to understand and interpret AI and ML algorithms become increasingly desirable. Visualization is a common tool to help users understand individual ML decision-making processes, but its use in demonstrating the global patterns and trends of a ML model has not been sufficiently explored. In this paper, we present a visualization technique using subspace projection to visualize ML models as scalar valued multi-dimensional functions to help users understand the global behaviors of the models in different 2D viewing spaces. A formal definition of the visualization problem will be given. The visualization technique is developed using an interpolation-based subspace morphing algorithm and a subspace sampling method to generate various renderings through projections and cross-sections of the model space as 3D surfaces or heatmap images. Compared to existin