First, consider these 2017 predictions in the IDC FutureScape on IoT. By 2019,
- At least 40 percent of IoT-created data will be stored, processed, analyzed and acted on close to, or at the edge of, the network.
- All effective IoT efforts will merge streaming analytics with machine learning.
Success in IoT applications requires an analytics-centric, data-driven culture. For example, when switching from scheduled maintenance of devices, machines, equipment – things – to predictive maintenance, actions are no longer determined by the calendar. They are determined from data flowing from the things by applying logic, decision rules and models. In other words, actions are determined through analytics.
Increasingly, analytics will be applied less in the datacenter or cloud and more at the edges and aggregation points of the network. Streaming data provides movement to the right place; streaming analytics provides insight at the right place and time.
An artificial intelligence system is a computerized system that makes a decision or performs a task that a human could carry out. AI today is a form of advanced analytics that relies on machine learning, optimization and deep learning.
If analytics is a requirement for success in IoT, is analytics in the form of AI a necessity for IoT to fulfill its potential? Is the artificial intelligence of things – AIoT – the ultimate success story of the Internet of Things?Interested in learning more about IoT? Visit us at IoT Solutions World Congress, Oct. 16-18 in Barcelona, Spain
From smart connected devices to artificially intelligent things
As noted in Harvard Business Review, a smart, connected device is made up of several elements:
- Physical elements, such as the mechanical and electrical parts.
- Smart elements, such as the sensors, processors, storage and software.
- Connectivity elements, such as ports, antennas, and protocols.
Physical components are amplified by the smart elements, which in turn are amplified by connectivity.
For a smart, connected thing to be a thing in the AIoT, it needs to make a decision or perform a task that a person could do. There are many smart devices that are not AI devices. Being controllable from an app or merely learning user preferences is not sufficient. A home heating system that learns temperature preferences is not an AI system unless it does something. It adjusts the temperature on your behalf. An autonomous vehicle is an AI system – it drives for you. When it is connected to other cars or the internet it is a thing in the AIoT.
Adjusting the blade angle of downstream wind turbines, rebooking a flight and hotel, making a recommendation, interacting with the physical world, recognizing faces, translating language or approving a loan are artificially intelligent feats when done by an algorithm. Imagine what could happen when these feats are done by systems connected to each other.
From collecting data to collective learning
Connectivity amplifies the smart elements of products and devices by externalizing their capabilities. It enables monitoring, control and optimization.
By itself, connecting things does not promote learning, but it paves the way. Many IoT applications rest on sending data to the cloud or datacenter, analyzing and modeling the data and applying the insights. They provide a result and possibly push modified logic back out to the devices.
The ability to adapt, to change behavior over time, is another keystone toward personalization and collective intelligence: We want devices to learn from their specific usage, and we want devices to learn from each other. We experience today the crudest form of collective intelligence: systems that are trained on information gathered from the collective, without much personalization. The generalizations of the trained system apply to us all. The same Siri lives on your phone and on mine, the same Alexa dwells in your home and in mine.
To promote learning and collective intelligence, the connected devices need to understand the value of information provided to them and apply them in informal, self-directed networks. The cameras stationed around the soccer pitch adjust aperture based on the local light conditions and know to ignore whether it is sunny or cloudy five miles away. The cars in the proximity of a highway accident communicate traffic density to route traffic to suggest detours. A device that dispenses asthma medication understands the needs of the individual patient and can adjust the medication based on information about smog conditions at the business traveler’s arrival destination.
The opportunity in the AIoT is to promote learning and personalization at the same time. We all want to be treated individually, with our habits, patterns and preferences considered. Two pieces of industrial equipment of the same make and model do not perform identically under different conditions and are likely not used identically. Treating them alike misses IoT opportunities for greater operational efficiency, greater safety and better use of resources.
These AIoT systems might well be operated by another AI system, for example, one built on reinforcement learning. It is unlikely that we can achieve the high degree of personalization without automated systems that are flexible enough to optimize themselves for thousands of scenarios. If we can train an algorithm to suggest the next best move in the complex game of Go, we can train an algorithm to adjust the chillers in a datacenter or the blade angle of a wind turbine, or to dispense the right amount of medicine at the right time.