III: Medium: Visually Interactive Neural Probabilistic Models of Language

点击翻译按钮获取中文摘要

The application of machine learning to automate everyday tasks is becoming increasingly common. While automation has the potential of yielding higher efficiency and improved outcomes, it can lead to unpredictable mistakes that can be hard to analyze and correct. Users of machine learning algorithms need better explanations for predictions and choices that are made. Moreover, to prevent harm, a user should be able to intervene in and control the decision process of the algorithm. This award is primarily concerned with applications in neural language models, i.e., machine learning systems that communicate using natural language. Because these systems interact with users using text or speech, it is essential to avoid misinformation from automated approaches and to retain human agency. Developing explainable and controllable artificial intelligence methods will empower users to collaborate with automation tools and gain efficiency and performance benefits while at the same time preventing harm and misinformation. This project targets the development of methods and visually interactive tools that allow researchers to develop, examine, and correct probabilistic neural models of language. Co-designing machine learning models and visual interfaces will be a necessary step towards interpretable models for common use-cases such as language summarization, translation, and data-to-text applications. To achieve these interactive and collaborative systems requires developing novel probabilistic neural network models with latent variables that can act as "hooks" within the model. These hooks correspond to interpretable decisions that a model has to take and that enable end-users to overwrite and interact with model decisions. In a second step, the project will develop query and visualization methods that utilize these hooks to allow users to explore, debug, and improve neural models on real examples through interactive user feedback. The project progress will be evaluated using quantitative methods from machine learning, qualitative and quantitative user studies, and long-term longitudinal observations of user engagement. The project will result in an extensible software framework for visually interactive analysis of neural sequence models that will assist other researchers and developers in their application domains.This award reflects NSF's statutory mission and has been deemed worthy of support through evaluation using the Foundation's intellectual merit and broader impacts review criteria.