Applying AI and ML to IoT-generated Data
Artificial intelligence (AI) and machine learning (ML) play a vital role in the future of the Internet of Things (IoT). At last week’s IoT World in Santa Clara, this was a major focus with a track dedicated solely to the topic. Sessions ranged from discussing the need for AI and ML to debating where to deploy AI and ML – at the edge or in the cloud.
The importance of AI and ML to realize the potential of the IoT
If you look at the current installed base of Internet connectable devices, IHS Markit is currently tracking 35 billion of them, and this is what comprises the “Internet of Things”. That is a lot of things, generating a lot of data, and the possibilities of how to use that data could be revolutionary. However, at present, the IoT is mostly point to point solutions, where a sensor on a water pump might tell you it’s close to empty and needs to be refilled soon, or a person might even be able to control that pump remotely, turning it on and off as needed. This point to point connectivity and communication is enough for applications that exist in silos.
However, when you look at the use cases for the IoT that excite the industry and drive investment from enterprises, it’s not about remote control; it’s about smart cities, autonomous driving and even smart retail. These industry-changing use cases require a variety of devices, across various applications, to connect and interact with each other. It’s not only just about sharing data, but having algorithms designed to identify patterns, flag anomalies, then make decisions and even act based on those anomalies, all without human intervention. That’s the future of the IoT and that’s all reliant on AI and ML.
The debate between the cloud and the edge
IoT World drew a variety of vendors with very different approaches to applying AI and ML to IoT-generated data. Some were providing gateways for close-to-the-edge intelligence, others were facilitating deployments of AI and ML in the cloud, and component manufacturers were demonstrating the capabilities of intelligence on the devices themselves. IHSM moderated a panel of experts from Actian, DeCypher DataLabs and Cirrus360 discussing this very topic.
The arguments for processing in the cloud are numerous, and primarily surround the massive data sets which leave few options but to utilize centralized data centers with their high-performance computing resources. Additionally, the availability of “off-the-shelf” AI and ML platforms being offered by major cloud providers draw in smaller enterprises not able to invest in the programming expertise required to build custom ML solutions. Perhaps, more importantly, processing in the cloud allows for a lower bill of materials (BOM) on the device itself, not requiring an expensive processor.
Counter points from edge proponents include the still-decreasing cost of compute, making AI-capable chips ever-less impactful to device margins. The biggest driver for edge intelligence is latency, with use cases like autonomous driving that have latency requirements of 2 milliseconds or less. In these mission critical cases, there isn’t time for data to be collected, travel to a data center, be analyzed, and then a decision made and returned. In these cases, on-device, high-processing capabilities are a must.
In the middle of these two extremes lie the world of gateways, which is a commonly utilized approach to managing data in many applications. They are also critical for extracting data from devices that cannot connect directly to the internet and cloud platforms, but are used widely used in commercial and industrial applications, such as Modbus, Bluetooth Low Energy and ZigBee. IHS Markit estimates well over three million cellular IoT gateways shipped in 2018. Gateways are a compromise for use cases that have compliance regulations or security concerns that require on premise compute. Their also a good solution for use cases with latency requirements that fall in the few to 50 millisecond range. Smart factories, with their cybersecurity concerns, are a prime market for gateways. Similarly, hospitals, which are aggregating data from thousands of devices at once and also have strict compliance requirements, are another case in which putting AI and ML on a gateway is a good option.
IHS Markit has summarized the requirements for AI and ML into five major components – each having an impact on the decision to deploy AI and ML in the cloud, on a gateway or at the edge. These five criteria are latency, privacy and security, power consumption, processing power, and cost of data communication. Lower latency requirements obviously drive you closer to the edge. Privacy and security can push you in either direction depending on the enterprises’ own cybersecurity and compliance capabilities. Energy consumption and processing power have an inverse relationship – meaning the greater processing power you require on the device, the more energy that will be required. Therefore, you wouldn’t want significant compute occurring on your smart watch, given the limited battery life. Lastly, cost of communication is a factor when thinking about placement of AI and ML, given that it costs money to transport data to and from devices, gateways, and the edge.
We are still at the early stages of AI and ML adoption in the IoT. However, at IoT World in Santa Clara, with industry experts gathered, it was clear that all are in agreement these tools are necessary to accomplish the vision of the IoT. Not only do we need it today, but when you consider the impact of 5G ten years from now, it’s a must. 5G will move us from the ability to connect 60,000 nodes per square kilometer with NB-IoT, to having one million nodes per square kilometer. That’s too much data to NOT have intelligence handling it.
Beyond the agreed upon role of AI and ML in the future of IoT, we face the question of where to deploy. IHSM believes that intelligence on-devices will certainly grow in the future, as cost of compute declines, and mission critical use cases abound. However, cloud platforms providing AI and ML to small and medium enterprises play an important role in wider spread adoption of the technology. In either scenario, cloud or edge, a gateway will likely be involved given its aggregation and translation capabilities between connectivity technologies. In summary, we expect to see a growth in all deployment types, with no one architecture perfect for all use cases.