Connected AI is More Than the Sum of its Parts

By Moshe Sheier

April 28, 2025

AI is everywhere and growing fast in the IoT (hence AIoT) since we realized that a cloud-based AI dependency is slow (latency), expensive (power-hungry servers, transmission cost), not always reliable (link downtime), and a privacy risk. Doing more of the AI processing locally on IoT devices solves these problems. Nevertheless, almost all applications depend on connectivity to be useful, because AIoT devices must still communicate their intelligently locally processed results to the outside world. If AI on the IoT device is doing more of the work, do we now only need a lightweight connection to handle the link back to the cloud or other devices? Not at all. Wireless connectivity technology is also getting smarter. Certainly in communication protocols, by leveraging AI to increase reliability and efficiency, conversely by adding new types of channel sensing, and by supporting distributed learning among a community of related AIoT devices. The link between AI and communication is not getting looser – it is getting stronger!

The Opportunity

Forecasts project that 6 billion IoT devices (based on TinyML chipsets) will ship in 2030, across many application domains from agriculture to smart homes, transportation, wearables and healthcare, smart cities and utilities. Depending on application these must support a range of communication protocols, all the way from cellular and Wi-Fi to Bluetooth and 802.15.4. It is fair to assume that anyone building an AIoT product will want to offer more than just table stakes – an AI function and a connectivity function side by side – if they can differentiate through enhanced leverage of an integrated solution. What might that look like?

AI Contributing to Communication

In crowded communication environments it becomes very important for an AIoT device to select the most available channel with the least interference, and to be able to alter this choice adaptively as ambient conditions and traffic demands change at and between access points or base stations. In simpler times this channel estimation/optimization was handled through precomputed lookup tables, but now AI management has become essential to keep up with these more complex demands.

This optimization is not only important for throughput. Safety-critical applications, such as automotive apps through V2X or in surgical and industrial robotics, all depend on ultra-low latency. This is a key component of the 5G cellular standard (and beyond) and requires guarantees from both network and end-user devices, increasingly served by AI-based channel optimization.

Another emerging application is positioning, especially valuable to locate moving devices (packages, shared bikes) in a smart city. Communication between base stations and edge devices can provide time-of-flight and angle of arrival data, though accuracy can be compromised by reflections and other factors. AI can mitigate these limitations through learning over time.

Communication Contributing to AI

This is a very exciting area, using communication (particularly Wi-Fi and cellular) as an additional sensing input to an AIoT device. By monitoring channel state information (CSI), commonly compromised by blockages and moving objects, then collating these inputs from multiple devices around a room/office/building/city, such a system can detect objects or people moving, even down to the level of detecting breathing rates.

This sensing input depends heavily on intelligent processing to separate interesting signals from noise, to eliminate reflections and don’t-care movement (such as overhead ceiling fans, or pets) from human activity. Applications extend from home security to gesture recognition, to non-intrusive health monitoring.

Exactly how much of a role AIoT plays in this application is still evolving but it is becoming clear, as in so many other applications, that some level of AI processing on edge devices – smart speakers, smart TVs, smart power sockets, etc. – will be essential given their natural distribution around the environment. Local intelligence is also essential to pre-process and reduce what must be sent to a central hub for final classification.

A different but equally interesting example can be found in federated learning. For fleets of systems – might be cars, or autonomous cars, or autonomous office cleaning devices – learning must be an ongoing process and can’t depend on shipping massive amounts of data through backhaul to a cloud-based training network. Training instead should start to be developed locally on each edge system, then that enhanced AI model can be shared between the local nodes and shipped back to a cloud-based training consolidator. This is the “federated” part of federated learning. Each edge system is responsible for contributing its own learning to the greater good. AI at the edge obviously plays a role, and so does communication because it must handle efficiently local training upload and eventually revised global training download.

Companies who can offer technology and expertise in both communications and AI are very rare. We have this expertise and technologies at Ceva and are accustomed to providing fully integrated solutions or options in which AI and communications sit side by side. If you would like a discussion on trends we are seeing and how we might be able to help, checkout our website and give us a call.