Can Android Pie use DeepMinds AI to extend battery life?

category:Internet click:556
 Can Android Pie use DeepMinds AI to extend battery life?


The following are the main contents of the article:

In January 2014, Google spent $400 million to acquire DeepMind, an artificial Intelligent Company based in London. At that time, it was not clear what Google and its parent company Alphabet would get from the deal. Four years later, the DeepMind team, which focused on developing AI for Google, began to pay dividends.

The launch of Googles latest mobile operating system, Android Pie, involves the launch of DeepMinds largest real-world machine learning system to date. Its AI has an ambitious goal. It hopes to solve one of the most frustrating pain points of modern smartphones: poor battery life.

Starting in the spring of 2017, DeepMinds London team began working with Google colleagues long before the developer preview of Android Pie (formerly codenamed P) was released. The result of cooperation is that they introduced two AI systems into the operating system. They are Adaptive Battery and Adaptive Brightness, which aim to prevent applications from consuming battery power in the background, and Adaptive Brightness, which aim to automatically adjust screen brightness to the environment in which the phone is located.

Ben Murdoch, an Android engineer, said the first batch of data from Android Pie developers, beta beta and generic versions showed that the system worked. He says applications running in the Android device backend wake up the central processing unit (CPU) 30 percent less frequently, and the amount of data transmitted through Wi-Fi and mobile signals in some cases 20 percent less. These two methods can reduce battery consumption.

We have seen a decrease in the variance we mentioned. Murdoch adds, Most users often feel that their mobile phone battery consumption is much faster than expected or normal. We call these unpredictable events battery bad days. We are controlling the bad days of batteries.

Early stage

Although AndroidPie has been launched publicly, these systems are still at a relatively early stage. The public version of the operating system has been available for download since August 6, but is still available only on a few phones. There are more than 2 billion devices running Android versions worldwide, but most of them run on older operating systems. The latest data released by Google before the launch of Pie showed that only 14% of the devices were in use of Oreo system. )

DeepMinds AI analyzes how users of Android devices use their applications. It has two layers, using the timestamp of the applications opening to predict the next time the application opens. The machine learning model learns the usage patterns of applications -- removing application names and details to prevent them from being treated unfairly -- and then predicts which applications are often used. Then, we get the possibility data that each application is divided. If two applications are used the same way, they may get the same prediction because they will have the same input data. But in fact, they may be two completely different applications, Gump said.

Applications placed in the upcoming group can run unrestricted, while those placed in the lower priority group are subject to different restrictions. Limits increase when applications begin to find themselves in groups that are about to be used, frequently used, or rarely used. Restrictions on those applications include: devices must be charged, or devices must be connected to the network, Murdoch said. Other restrictions include stopping the application setting alarm clock and waking up the phones function. Applications can also be limited to responding to messages received through the cloud, and the background and network activities of those applications in rarely used groups are completely limited.

This may have an impact on user experience. When Android Pie users open the adaptive battery system, the system issues a warning: Notification alerts may be delayed. So if you dont use Facebook very often on your phone, you may delay receiving push notifications from the application. Applications are scanned hourly to predict their usage, and AI processing is done on all devices.

AI is extremely complex on mobile phones.

Previously, DeepMind pushed its AI technology to Googles data center. Its machine learning directly controls the cooling of giant buildings filled with servers and Internet infrastructure, which companies claim save energy. However, getting involved in mobile phones is another matter.

The application of machine learning technology on mobile devices is extremely complicated. Gamble said. Although mobile phones are more powerful than ever, their computing power is still far below that of larger systems, which rely on more resources to process data. One thing we are sure is that the first iteration of the model is very computationally intensive. Gamble added. This is especially important for non high-end phones. Adaptive batteries and adaptive brightness systems were initially tested on Googles Pixel smartphone, but as it came out of the prototype stage, it was extended to other phones.

As the machine learning model becomes more and more widely used in the real world, Android and DeepMind have been able to upgrade the AI before the advent of Android Q in 2019 in order to solve any problems in the model in time. These models are built and deployed on their own Android APK, and we can flexibly update them as needed through the PlayStore application store, Murdoch said. Google can push updates to machine learning whenever it feels like it has made enough improvements. One of the things we intend to do over time is monitor the performance of these models in this area and adjust them as needed. At present, it is impossible to fully explain the effectiveness of machine learning technology, and it is not known whether there will be significant improvements in cell phone battery life. If there is a problem with AI, people will naturally complain. Murdoch said, ideally, users do not find problems, which is our greatest success. (ROBOM) source: NetEase science and technology report editor: Yao Liwei _NT6056

As the machine learning model becomes more and more widely used in the real world, Android and DeepMind have been able to upgrade the AI before the advent of Android Q in 2019 in order to solve any problems in the model in time. These models are built and deployed on their own Android APK, and we can flexibly update them as needed through the PlayStore application store, Murdoch said. Google can push updates to machine learning whenever it feels like it has made enough improvements. One of the things we intend to do over time is monitor the performance of these models in this area and adjust them as needed.

At present, it is impossible to fully explain the effectiveness of machine learning technology, and it is not known whether there will be significant improvements in cell phone battery life. If there is a problem with AI, people will naturally complain. Murdoch said, ideally, users do not find problems, which is our greatest success. (ROBOM)