The recent boom in artificial intelligence (AI) benefits from the open collaboration of the open source software (OSS) community. An increasing number of OSS developers are contributing to AI projects by submitting pull requests (PRs). However, the PR quality submitted by external contributors varies, and the AI project management teams have to review PRs and ask contributors to revise them if necessary. Since the revision exerts a direct impact on the review efficiency and acceptance of PRs, it is important to achieve a better understanding of PR revisions. This study conducts an empirical study based on a set of PRs and their revision histories collected from the TensorFlow project. It first manually analyzes a sample of commit messages, reviews PR comments, and constructs a taxonomy of revision types. Then, according to the defined taxonomy, a large set of PR revisions are manually labeled. Based on the dataset, the frequency and order of each type of revision are explored. Additionally, this study also investigates the frequency distribution, order distribution, and correlation relationship between different types of revisions. The empirical findings show that there are 11 different types of revisions which can be classified into three categories. Evolvability revisions occur more frequently than other revision types, and functional revisions are more likely to occur in the early PR updates than evolvability revisions and other types of revisions. Structure-related revisions have a high chance to co-occur or adj-occur with other revisions. Configuration-related revisions or rebasing revisions are more likely to appear in succession. The empirical results can help AI open source practitioners and researchers better understand the PR revision process, especially guide the review and revision behaviors of PRs and improve the collaborative efficiency of open source groups.