Click here to learn more about author James Kobielus.
Machine learning is beginning to sink roots into every node of the Internet of Things (IoT). In this new era IoT endpoints will increasingly be powered by statistically driven algorithms that process real-time sensor data and drive autonomous decisions right then and there. I recently published a blog on DATAVERSITY.net on this very topic, which I urge you to read as you process the following remarks.
As this trend intensifies, dumb sensors will become the exception rather than the rule at the edges of the IoT’s fog-computing fabric. Going forward, developers will increasingly embed algorithmic microservices that enable increasingly resource-constrained IoT endpoints to handle most of this processing locally and more rapidly and flexibly than any cloud service. In other words, most IoT edges will have the ability to make decisions and take actions autonomously based on algorithmic sensing of patterns in locally acquired sensor data. These embedded capabilities will range across the entire spectrum of algorithmic approaches that we associate with machine learning, deep learning, and the like.
For those who wish to get steeped in the new paradigm of fog computing at the IoT’s edge, I recommend Joe McKendrick’s recent article discussing the new OpenFog Consortium’s reference framework. Key architectural principles in this framework include scalability, openness, autonomy, agility, programmability, and hierarchy—as well as the need for pervasive algorithmic intelligence all the way out to edge devices. It’s an excellent architectural overview for anyone trying to understand how all functions, not just machine learning models, will evolve in an environment where more functions are embedded in edge devices.
However, the OpenFog Consortium hasn’t produced a specific discussion of how machine intelligence is likely to evolve as it moves to the edge of the IoT. With that in mind, here are my thoughts on how analytic algorithms will be fitted to their pivotal roles in IoT endpoints. As compared to analytic algorithms that execute on IoT/fog resources in the cloud, those on the edge will:
- Process datasets that are more locally-acquired, smaller in volume, lower in latencies, more specialized, and predominantly persisted in memory, as befits algorithms that are configured with sensors;
- Grow simpler and more specialized, in keeping with the task-specific nature of most IoT endpoints;
- Require less processor parallelism, consistent with the narrower workloads of specialized edge devices;
- Execute at higher computer-per-datum rates, owing to the hierarchical processing requirements of deep learning for tasks such as image, video, audio, and other complex pattern-recognition tasks;
- Be configured into simpler ensemble models, considering that specific machine/deep-learning algorithms will prove optimized for the prediction requirements of specialized edge devices;
- Be packaged into commoditized hardware and be flash-upgradeable with revised algorithms over wireless connections, in keeping with their requirement to run the latest production algorithms that have been trained on shared IoT/fog clusters;
- Express simpler feature spaces, consistent with the discrete, repeatable, structured, and specialized nature of the machine data processed by any specific edge device; and
- Engage in less interprocessor communication and infrastructure roundtripping, in keeping with the need for edge devices to operate in intermittently connected, low-bandwidth, autonomous-decisioning scenarios.
Embedded analytics will be as diverse as the IoT endpoints whose automated behaviors they drive. Endpoints will possess the different types of intelligence suited to their myriad applications, tasks, and deployment models. And, let’s face it, not every endpoint will be able to make decisions and take actions autonomously.
Under any plausible fog-computing scenario, some IoT sensors will still be as dumb as dumb can be.