Artificial intelligence (AI) and machine learning (ML) have been hailed as significant opportunities for mobile operators. AI/ML has the potential to revolutionize the way operators plan and manage their networks, significantly contributing to the introduction of what Deutsche Telekom calls “brutal automation”, and to better detection of security threats.
These techniques could also help MNOs rethink their user experience, supporting new interfaces for consumer applications, Alexa-style, and new customer service tools like chatbots.
Another opportunity lies in the assumption that all companies delivering AI-enabled services to mobile users will need to do so over the cellular network. AI and ML have the potential to make sense of the vast quantities of data generated by connected people, cars and ‘things’, and to transmit instructions back to those objects so that they can act autonomously. Those connected robots, cars and sensors will have to communicate with the central AI engine in the cloud, often over mobile links. This will boost traffic, and where that traffic is critical (like intelligent driving instructions issued from the cloud), there could be the chance for MNOs to charge premium rates to B2B customers. As network slicing emerges, it is easy to see the potential for slices which are optimized for AI-driven applications such as predictive maintenance or drone control.
But there is a risk that AI, far from transforming the MNOs’ model, may reduce the value of their networks.
Many of the most promising services require very low latency and very high QoS, and so there is a move to push processing resource to the edge, especially for the most mobile applications, like smart driving (static applications, such as home digital assistants, are likely to use WiFi so these mobile use cases are vital to the MNO). And as more and more AI capability is shrunk down to run on small devices, rather than supercomputers, the requirement to transfer data and instructions constantly between the device and the cloud will be reduced, along with the MNO’s place in the value chain.
Some AI platforms are adopting a form of ‘batch processing’, with a high percentage of activity going on at the edge and selected results uploaded to the central cloud periodically to enrich the machine learning base.
And some AI-related activities emerging from university labs indicate how the industry might look beyond the batch idea to eliminating the need for wide area networks altogether. For instance, researchers at the University of Waterloo in Ontario, Canada say they are working on platforms that allow AI and neural networks to function free from the cloud and the internet, because they can adapt to the removal of compute power and memory. The scientists, in effect, teach AI that it can perform without hefty resources, the benefits being lower data costs, higher responsiveness, better privacy and the ability to deploy AI-driven applications in remote areas.
In the research, the group has placed the neural network in a virtual environment and “progressively and repeatedly deprived it of resources”. The researchers reported recently that “the deep learning AI responds by adapting and changing itself to keep functioning”, becoming progressively smaller so that it is able to survive.
The researchers say they have been able to reduce the size of deep learning AI software for object recognition, a key function for intelligent and autonomous driving among other applications, by 200 times.
Mohammad Javad Shafiee, a professor at Waterloo and the system’s co-creator, says the work will boost current efforts to run deep learning engines on small devices such as smartphones or robots, allowing them to run even when they are not connected to the cellular network.
“When put on a chip and embedded in a smartphone, such compact AI could run a speech-activated virtual assistant and other intelligent features,” he told local news reporters.
This is just the latest example of technology projects to shrink AI onto mobile devices and therefore reduce reliance on long distance connectivity. Intel and Qualcomm have been locked in combat over this emerging opportunity and this week’s Consumer Electronics Show (CES) in Las Vegas saw Huawei and Samsung upping their game too.
Last fall, Qualcomm made its Snapdragon Neural Processing Engine software development kit available, providing programmers with tools to create on-device neural network-driven applications. Like the Waterloo team, Qualcomm argues that carrying out most of the AI processing on the device can improve reliability, privacy protection and bandwidth efficiency, compared to solutions which have to transmit data to the cloud for processing.
“We started fundamental research a decade ago, and our current products now support many AI use cases from computer vision and natural language processing to malware detection on a variety of devices — such as smartphones and cars — and we are researching broader topics, such as AI for wireless connectivity, power management and photography,” said Matt Grob, EVP of technology at Qualcomm.
Meanwhile, Intel has been harnessing its acquisition of machine vision processor start-up Movidius to push AI to the edge. Movidius’s low power Myriad X vision processing unit (VPU) claims to be the first system-on-chip (SoC) with a dedicated Neural Compute Engine. It is “specifically designed to run deep neural networks at high speed and low power without compromising accuracy, enabling devices to see, understand and respond to their environments in real time,” said Intel.
Intel Movidius exec Remi El-Ouazzane said: “With this faster, more pervasive intelligence embedded directly into devices, the potential to make our world safer, more productive and more personal is limitless.”
Also powered by Movidius is Intel’s Neural Compute Stick, a plug-and-play device which sells for $100 and does not need the cloud. It is designed for prototyping and deploying neural vision networks at the edge with no internet required and is the size of a PC memory stick.
At CES, Samsung unveiled its latest Qualcomm challenger, the Exynos 9 Series 9810 SoC for smartphones and other mobile devices, with localized AI firmly at the center of its specifications. The product boasts deep learning processing functionality as well as a 2.9 GHz custom CPU and an LTE modem with six-way carrier aggregation support to enable 1.2Gbps download speeds.
“The Exynos 9 Series 9810 is our most innovative mobile processor yet, with our third-generation custom CPU, ultrafast gigabit LTE modem and deep learning-enhanced image processing,” said Ben Hur, VP of system LSI marketing at Samsung. “The Exynos 9810 will be a key catalyst for innovation in smart platforms such as smartphones, personal computing and automotive for the upcoming AI era.”
The SoC is likely to turn up in at least models of the next Galaxy S smartphone generation, improving consumer apps that rely on capabilities such as facial recognition and object detection. Driving AI into the heart of the mobile consumer experience is also a priority for Huawei in its latest partnership, with Chinese internet giant Baidu. The companies are developing an open AI platform to help developers create new mobile services, and they will also cooperate on enhanced voice and image recognition technology for use in Huawei’s smartphones.
In a joint statement, the companies said their work would “lead the new era of mobile AI … It will help lay the foundation for a sustainable mobile and AI ecosystem, so that future technology can better understand users, better serve people, and promote better economic and social outcomes around the world.”
“The future is all about smart devices that will actively serve us, not just respond to what we tell them to do,” Richard Yu, CEO of Huawei consumer business group said in a statement.
So with all this activity at the edge, do consumer and B2B AI-enabled services really promise to enhance the MNOs’ revenues as they move towards 5G. It is a critical question. One of the chief justifications for investment in 5G, for many operators, is the ability to support the very low latency and very high data rates required by machine learning, and associated functions like augmented reality. If the majority of the heavy lifting is to be done at the very edge, a crucial aspect of 5G return on investment may be lost.
This makes it important for mobile operators to be able to monetize AI even when it is at the edge, perhaps by supporting shorter range links (from a gateway to a device perhaps, or between an autonomous car and roadside infrastructure). But the closer to the edge the value lies, the more the revenues and the user experience will be controlled by others, such as Google and Amazon. The MNO’s power lies in its exclusive ability to support long-range, high quality mobile links to the cloud – hence the ambivalence that many operators show about their ability to profit from deploying edge compute platforms like ETSI MEC (Multi-access Edge Computing).
The right balance between the economies of scale and efficiency of a centralized, cloud-based system, and the responsiveness of a distributed, edge-based approach, is one of the key decisions facing enterprises and service providers – and that, in turn, will affect how mobile operators need to plan their networks.
This is not just about consumer services such as digital assistants and connected cars. It is also about addressing the far bigger opportunity for new revenues, presented by B2B customers’ rising use of AI in mobile environments. But industrial and enterprise organizations are also wanting to push more intelligence to the edge. In a recent interview with EETimes, Alex Tepper, founder and head of corporate and business development at Avitas Systems (a GE venture to apply AI in the industrial sector), said that the key limitation of AI today is its ability to compute on the edge device itself – in a drone, for instance, enabling it to change its behavior in-flight without the delay inherent in receiving instructions from the cloud.
So the critical architecture decision that every MNO needs to make in the next few months or years is how to balance resources between the center and the edge of the network (see Wireless Watch August 4 2017 for our analysis of this). And in doing so, the most critical business decision is how to monetize a network in which the long distance connections to the cloud are less valuable than had been anticipated.
Of course, telcos will be using AI themselves, to help plan and optimize their networks more intelligently and dynamically, and map network investment to revenue value in the 5G era. Once networks are virtualized, the potential to transform their efficiencies and responsiveness should be limitless. But only if the planning of the physical, as well as the virtual, resources is fit for purpose. Depending on the use case and business model, there should be an optimal allocation of network and compute resources to different places along the chain, from cloud to switch to local gateway to the device itself. Ideally this should be capable of adjusting to changing circumstances in terms of traffic levels and type.
The ability of AI to help make these decisions is key – building on deep machine-based analysis of traffic patterns, user behaviour and history, and many other factors. A study of 48 mobile operators worldwide by Rethink Technology Research, conducted in autumn 2017, found that 58% were engaged in AI tests, trials or real world deployments, with machine learning being the most commonly used technology.
A newly published survey of the broader telco space by CapGemini concluded that, while telcos used to be AI laggards, they are now leading large-scale deployments, with 49% deploying the technology, ahead of an average of 36% across all industries. These are being led initially, in the majority of cases, by customer service applications, but a full 93% of telco adopters said they expected AI to increase efficiency and effectiveness, while 79% claimed to have seen a 10% boost in sales thanks to AI.
Vendors are starting to support the optimization aim. Nokia has made several announcements in recent months, including its Autonomous Care offerings, unveiled in May. Earlier this month, ZTE said MNOs needed to accelerate their network AI efforts and announced a roadmap towards an end-to-end platform which covers a wide range of telco-specific use cases from intelligent automated networks to new consumer services, and which incorporates the algorithms along with the chips and terminal hardware. The elements promised include ‘self-researching AI chips’, robot modules and intelligent terminals such as smartphones and smart home controllers.
“Complemented with high computing power, precision algorithm and data analytics capability, AI technology will lead to the evolution of highly intelligent autonomous, automatic, self-optimizing and self-healing networks,” ZTE said in its release. “At this stage, operators and vendors are still proactively exploring and seeking more efficient, stable and accurate AI algorithms and solutions to reduce the operation labor cost and effectively improve operating income. The platform can help operators introduce new technologies and build next generation intelligent network more conveniently amidst the ongoing advancement of AI technologies.”
ZTE’s inclusion of smartphone and controller devices in its AI portfolio indicates that the algorithms – which, before the days of cheap mass storage and compute power – required a supercomputer to run, can now be applied to a mobile gadget.
Telcos, then, have their own dilemmas about how far to push AI processing to the edge, in order to support their internal efficiencies. Edge-based AI improves responsiveness but an efficient way of updating the central platform is essential to avoid fragmentation. There are daunting issues of supporting smooth roaming for users who move from one AI-optimized, context aware cell to another with no such user experience. For challenges such as load balancing across different locations and times of day, a common view of the whole network is essential.
So the ability to do more AI at the edge does not answer all the questions of how to harness resources most efficiently as Tepper from Avitas makes clear. Avitas recently announced an alliance with Nvidia to work on enabling AI in inspection services for the oil, gas and transportation industries. Nvidia wrote in a recent blog post: “How do you send a human being to inspect a petroleum refinery flare stack — one that operates at hundreds of degrees and requires negotiating a high risk vertical climb? The answer is you don’t.”
While climbing a cell tower does not carry this level of risk, MNOs such as AT&T and T-Mobile USA have already experimented with drones to inspect and even install equipment, to save cost and liability. However virtualized, there will always be physical elements to a mobile network, and civil works can be the most expensive aspect of a roll-out, especially when it comes to large numbers of small cells to support urban densification.
Tepper points out that AI can create 3D models of an asset such as a cell tower, then layer “points of interest” on top of that to enable drones to spot problems and automate defect detection. Avitas and Nvidia are currently using truck-based AI engines to get closer to towers and industrial sites, but Tepper wants to get that intelligence into the drone itself. This is a far more complex, resource-hungry and mission critical task than supporting consumer applications based on vision processing, for instance, on a handset.
AT&T is also working on an edge computing model with AI elements to boost automation, revealed Marachel Knight, SVP of wireless network architecture, at the Mobile Future Forward conference in August. It aims to design its 5G RANs so that network computing components are geographically close to a tower or small cell to lower latency. It has already said that it plans to fit its edge computing platforms with high end GPUs and CPUs, and coordinate and manage all these elements with its software-defined network (SDN) controllers.
The goal is the same for AT&T and for GE (and many others) – to make AI highly personal and context aware, in order to go beyond automation and improved decision support, and enable new ways of working. On that journey, the right decisions, about how much to distribute or centralize, will help decide where the MNO fits into the complex AI value chain.