Visualizing Learned Models and Data for Exploratory Machine Learning

点击翻译按钮获取中文摘要

Visualizing Machine Learning Models and Data Choosing an appropriate model for the data or changing an existing model are both essential, but difficult steps in machine learning. Surprisingly, the trend in machine learning research has been to ignore that wonderful visual information processor--the human--and to build learning systems that are stand-alone and fully automated. This research is attempting to show that machine learning is a task best shared by humans and machine because of their unique capabilities. Human vision gives us built-in features such as motion detection and direction, stereoscopic depth, edge and shape detection, grouping by color, light, and shade. Because of these features, humans are good at quickly recognizing complex patterns in visual data, quickly detecting outliers in visual data, and visually manipulating a model to reflect the data. In contrast, machines are good at fast, accurate, and repetitive calculations necessary for machine learning. Not only is an appropriate division of labor between human and machine important, but a visualization of a learned model can serve as a visual explanation of why the learned model fits the data. Finally, visualization together with direct manipulation of the model can make it easier for the user to change the model to reflect the data, immediately see the results, and focus on interesting data regions. The impact of this research is that is may become easier to find good learning models and to visually understand why they are good.