Deploying AI Across Mobile and IoT
If AI can be made to work in low-power scenarios, smart devices will get a whole lot smarter.
Artificial intelligence (AI) is a powerful technology paradigm. Machine learning, deep learning and other AI-related technologies offer potentially vast business value. But powerful tech often requires powerful computing infrastructure, so AI is usually deployed on centralized servers and mainframes. To get maximum value from this technology, organizations will increasingly need to deploy AI in low-power scenarios—on mobile phones, tablets and especially across the billions of smart devices that make up the Internet of Things (IoT).
So, how can we deploy the power of AI in low-power scenarios?
IoT Driving AI Growth
If you follow tech trends, you’ll be only too well aware of the big, big future predicted for IoT. You’re probably accustomed to reading extraordinary projections estimating how many billions of smart devices will be populating the IoT in coming years. What you may have missed is a crucial element of the long-term vison for IoT—to make smart devices smarter via AI-enhanced services. This means that, when these smart devices become truly ubiquitous, AI will be everywhere. Literally.
Current technical limitations mean that smart devices aren’t really that smart. They’re essentially dummy units, which rely on centralized services to make intelligent decisions and coordinate with other phones, tablets, connected cars, smart meters etc. Some of these devices can monitor the environment or make changes when ordered to do so, but they seldom make decisions by themselves. We can expect this to change. Decision-making and orchestration will increasingly be pushed down to the devices themselves.
Three main developments will enable this change. First, with transfer learning, deep learning (DL) won’t require as much data as previously assumed, so it won’t be necessary to centralize data gathering. Second, hardware innovations will make it significantly easier for low-power devices to run complex DL systems. Finally, with the projected size of IoT, the current centralized approach won’t be sustainable, and it will be necessary to push computations down to the device level.
Before the benefits of these developments can be fully realized, though, there are a couple of significant issues that will need to be addressed.
Can AI be Agile Enough for IoT?
Currently, it seems that many organizations are struggling to generate value from their AI efforts. A recent study from Narrative Science found that, in 2016-2017, 31 percent of businesses that had adopted AI were either not tracking ROI or were not seeing returns on their AI investments. One factor probably contributing quite significantly to this shortfall in ROI is the lack of “agility” inherent in some of the technologies most vital to AI—specifically machine learning and deep learning.
With an average time-to-outcome of 52 days, machine learning projects are hardly suitable for agile software development cycles. And deep-learning applications are generally highly device-specific—so deploying DL across a range of heterogeneous devices will inevitably prove time-consuming. In this context, enterprises will need to adopt tools designed to bring agility to AI development. Right now, these tools don’t exist, but we can expect them to emerge over the coming years.
Security Essentials for IoT Success >
IoT + AI = Security Headaches
One of the big issues that is often raised around IoT is security. Events ranging from the seemingly benign (Burger King’s “okay Google” ad campaign) to the potentially catastrophic (Mirai-based botnet attacks) have brought home the fact that we are now surrounded by connected devices for which security measures are often all-but non-existent. The arrival of device-native AI applications will add even more complexity to the already imposing question of IoT security.
Specifically, the issue of data perturbation is likely to be significant. This refers to the practice of introducing “noise” into a data set in a way that may be invisible to the human eye, but which could nevertheless affect the decisions an AI makes on the basis of the data. In the context of IoT, this is likely to be a highly-significant threat for devices that make access decisions based on image recognition. But it could also be used to manipulate deep learning data to change a device’s behavior over the longer term.
Deep learning in connected devices is also likely to raise privacy issues. If hackers are able to retrieve learning data from a device, the personal information of anyone who interacts with that device could be compromised. This is a concern not only because of the insecurity of many IoT devices but also because DL architectures generally sacrifice security in favor of accuracy. Therefore, the migration of AI to low-power devices will also be dependent on the emergence of technologies that address the inherent security weaknesses if both IoT and DL.
While much progress is being made towards addressing the challenges associated with deploying AI directly to IoT and mobile devices, there is still plenty of work to be done. The long-term vision being pursued by IoT industry leaders makes this work not just desirable but essential. It will be interesting to see how the key challenges are addressed and what role the resulting solutions will play in making truly smart (and secure) connected devices both ubiquitous and beneficial to enterprises and their customers alike.