Google has developed a new model for training artificial intelligence (AI) that can train and improve the AI ​​algorithm directly on the user's smartphone. When large technology companies use machine learning to improve software, the process is usually very centralized. For example, companies such as Google and Apple collect information about how users use their applications, store this data somewhere on the server, and then use aggregated data to train new algorithms. Finally, users will get improved application updates. This AI algorithm training method is effective, but the process of updating applications and collecting feedback data is very time consuming. Moreover, this approach is not conducive to protecting user privacy because companies must store data on their servers about how users use their applications. In order to solve these problems, Google is trying a new AI training method and calling it Federated Learning. Federated Learning's training of AI algorithms is done directly on the user's device, rather than collecting user data somewhere on the Google server and using that data to train the algorithm. In other words, Federated Learning uses the CPU of the user's mobile phone to help train Google's AI algorithm. Currently, Google is testing Federated Learning on Android platform keyboard application Gboard. When Gboard displays the recommended search item according to the information input by the user, Gboard will remember the search item clicked by the user and the ignored search item, and then personalize the algorithm directly on the user's mobile phone. (For this test, Google has integrated a streamlined version of its machine learning software, TensorFlow, into its Gboard application.) These improvements will be sent back to Google and then aggregated by Google and published to all users. Google explained in a blog post that this AI training method has many benefits. First of all, it is more conducive to protecting user privacy because its training process is performed directly on the user's device and does not store user's data. Second, this training approach will allow users to immediately benefit from the personalized improvements of the AI ​​algorithm without having to wait for Google to release new application updates. Google said that the entire Federated Learning system has been streamlined and will not affect the battery life or performance of the user’s mobile phone. The training process will only be “in idle state and connected to power†and “free Wi-Fi accessâ€. "It will only happen.