The GSMA Terminal Steering Group set up a work item on AI mobile device guidelines in late 2018 to embrace the new development trend.
We have witnessed the high-speed development of mobile internet due to the fast penetration of smartphones, but after the big leap from feature phones to smartphones the industry has gradually entered a mature stage, with limited innovation and growth. Global smartphone shipments have declined and the usage amount and usage pattern of mobile applications has reached a steady state, remaining almost the same year after year. To break through and make further progress, the industry is in urgent need of a new innovation power and growth point.
AI is deemed to be the new force with promising prospect.
AI is not a new concept; it first came out in 1956 and from then it has roughly gone through three stages. Right now it is in stage three, which is marked by deep learning. Compared to the previous generations of AI technology, deep learning may reduce many manual interventions and have the intelligence of autonomy.
2006 is considered to be the year one of deep learning, but due its complicated and demanding nature it has only been implemented on mobile devices in the past few years. The device chipset that supports deep learning, mainly referring to deep learning inference, began to emerge in 2017, as well as the development of the software platform on mobile devices that supports it (either the OS or the chipset vendor’s private solutions). The frameworks and the light weight models appropriate for using on mobile devices were not ready until more recently. By the end of 2017, the industry value chain of on-device deep learning was almost formed, and there has been a visible boost in products (chipsets and devices) and market publicity ever since.
This wave of AI is marked by deep learning. Due to advancements in technology and engineering, on–device AI (e.g. on-device deep learning) has become a practical reality. On–device deep learning reflects the current level of technology and engineering: and when we talk about an AI mobile device at this stage it should be linked with deep learning and on-device AI.
But something unexpected has happened.
Under the hype of the media, the cognitive rate of AI mobile devices has increased rapidly. AI has become a key factor when making a purchase decision; according to an investigation conducted by China Telecom around 40% of people preferred AI mobile phones when they considered buying their next phone. In pursuit of market interests many device manufacturers have advertised their products as AI mobile phones and many applications have claimed to be AI applications overnight. This has lead to on-cloud AI and on-device AI mixing together, the old generation AI blurring the new generation. The functionalities and performances are uneven, and there is a big gap between the bad and the good: the chipset functionalities and performances vary a lot, and the AI application performances are very different. The abuse of the concept of AI mobile devices, the inconsistent functionality, and the varying performance highly confuses and frustrates end users.
TSG believes it is in urgent need to provide industry guidelines to help facilitate the benign development of AI mobile devices, and to ensure and enhance the end user’s experiences. This is the starting point. The necessity of putting forward AI mobile device guidelines is not only for the aforementioned reasons but, most of all, it is because of the important implications of an AI mobile device.
AI mobile device is a revolution. On-device AI will give phones the ability to learn from the environment and react autonomously based on the algorithms. This agent like “learn and react” characteristic distinguishes and differentiates AI mobile devices from smartphones. This brand new characteristic will stimulate the growth of new applications. From a historical point of view the AI mobile device is a revolution. The significance of this revolution is no less than that of the smartphone. Having more autonomous apps on mobile devices will be the trend.
AI mobile device enables better privacy preserving and security. We’ve all started to become increasingly aware of how big a concern data privacy is. However the expectations of technology’s ability to personalise apps and services to meet people’s specific interests, location and context continue to grow. Enabling this customisation requires the use of, and access to, some level of personal data (usage patterns, etc), which is where the dilemma lies. However, with AI-based software and hardware capabilities becoming available on mobile devices, more of the data analysis work can be done directly on devices. This means someone could get the same level of customisation and personalisation they’ve become accustomed to without having to share their data externally with the cloud. AI mobile devices make it possible to better protect privacy, and say no to sensitive data collection.
AI mobile device will drive the network to change. Mobile devices are battery constrained, so it is reasonable to offload computation tasks to MEC or the cloud in case the delay is tolerant. AI model’s training are usually conducted in the cloud. Consequently the cloud and MEC will need to evolve to better support AI computation.
Moreover, the legacy network is a communication centric network overlaid with cloud and MEC. This “transmit and then compute” model is not efficient. Academia is investigating a new type of network in which transmission and computation are not separated, instead being carried out at the same time, and thus radically changing the air interface and radio network.
Based on this, TSG’s AI mobile device guidelines will define the concept of an AI mobile device, and clarify what an AI mobile device is in the context of current technology and engineering. It will define the basic requirements for hardware and software and will also define the basic performance requirements for frequently used AI applications, aiming to align the industry. Furthermore, the AI mobile device guidelines will tackle the other key issues of far-reaching significance: privacy and security, network requirements, etc.