Huawei and Apple Make On-Device Machine Learning a Reality in 2017

So much of the news today about artificial intelligence (AI) has to do with intelligent systems beating humans at games or the potential for robots to take human jobs. Very little of the news is about the use of AI technologies, including machine learning (ML) and deep learning (DL) that is commercially available in mobile devices such as smartphones. More importantly, how quickly the use of AI is moving from the cloud to being processed or trained on the device itself.

Huawei was the first mobile device company to release an AI-powered smartphone that performs ML without relying on the cloud when it announced the P10 in February 2017. Powered by a Kirin 960 SoC made by Huawei subsidiary HiSilicon, the handset uses ML for two on-device activities: Optimizing device memory while the phone is charging; and predicting the user’s next touch-screen location to fetch and preload app content. Both uses benefit the overall user experience by providing faster device responsiveness.

Apple unleashed its next round of mobile devices during a September 2017 launch event. The Apple iPhone 8/8+ and iPhone X utilize an Apple-developed processor chipset dubbed “A11 bionic”, which contains ML processor cores in addition to the traditional CPU and GPU cores. The premier use for ML in these smartphones is for facial recognition in addition to authentication (called FaceID) for the iPhone X. Both iPhone families will also use the recognition engine to power animated emojis.

Announced in October, the Huawei Mate 10 is the follow-on generation of AI-powered smartphones from China’s handset leader powered by its latest Kirin 970 chipset, which has the addition of an integrated neural network processing unit (NPU).

Huawei truly surprised everyone by being first-to-market with a mobile device capable of on-device learning. While the initial handset application is utilitarian, working in the background while the device is charging and idle, it strives to provide a better user experience over the lifespan of the device. Could you imagine this technique being applied to other devices, such as automobiles? While parked, your vehicle is optimizing itself for safer and more efficient future trips. That’s amazing and, in the next couple of years, ABI Research expects most smartphones will have this AI-enabled capability along with initial market entry into wearables, smart home hubs, and driverless vehicles.

Device-powered intelligence at the network edge starts with enhancing the processor chipsets. Several companies have announced plans to support on-device AI-powered capabilities during 2017, such that developers can start trying out innovative ideas. 

The launch of a processor with dedicated machine learning functionality is significant for the smartphone market and for the adoption of AI technologies on mobile devices. Huawei was one of the first chipset manufacturers to pursue dedicated AI processing for a mobile device. Qualcomm also touts machine learning in its Snapdragon 835 chipset along with a Neural Processing Engine SDK for the Snapdragon 820 series processor.

The majority of machine learning on smartphones today is done using the cloud. Voice assistant applications are a good example of this. Using the cloud to perform processing, data set training, and storage does not require any resources from the installed base of smartphones while providing convenient and productive ways for people to interact with their device. It does, however, require a high-speed data connection to deliver a great user experience. Enabling the device to perform training and processing regardless of the mobile broadband data connection offers new application possibilities.

ABI Research forecasts that more than 675 million smartphones will be used for cloud-based machine learning applications, led by voice assistants, in 2017.

Announcements and commercial availability of products are a leading indicator that new generations of smartphones will do machine learning on-device. The benefits of on-device learning include the responsiveness as well as the personalized nature of the results (using data gathered by a specific device instead of using a large base of smartphones).

While the specific application benefits will not be fully understood until chipsets are in the hands of application developers, one could see augmented reality, enhancing video editing, faster object recognition in photos, providing better user authentication, and optimizing the performance of the device as examples of on-device machine learning advantages.

ABI Research further estimates that only 34 million smartphones will be used for on-device machine learning applications in 2017, a reflection of the early stage of today’s market.

Additional ABI Research content that is useful in furthering understanding of the AI and Machine Learning landscape, includes:

Artificial Intelligence & Machine Learning: Foundational Concepts and MD-AIML-101: Artificial Intelligence & Machine Learning