Edge AI refers to deploying the computational and processing capabilities of artificial intelligence close to the source of data generation, that is, on “edge” devices, rather than relying entirely on remote cloud data centers. In simple terms, it enables devices themselves to run AI models and make real-time intelligent decisions. This article summarizes the advantages and necessity of Edge AI based on the book “AI at the Edge.”
Bandwidth
Typically, the data captured by IoT devices far exceeds their available bandwidth for transmission. This means that the vast majority of sensor data captured is not even utilized and is simply discarded.
Smart sensors can monitor the vibrations of industrial machines to determine if they are operating normally. They can use simple threshold algorithms to analyze when a machine is vibrating too much or too little, and then relay the results via wireless connection. However, if patterns in the data can be identified, it can provide predictions of impending machine failures. If we had ample bandwidth, we could send sensor data to the cloud for some analysis to achieve failure predictions. However, in many practical applications, there is not enough bandwidth or energy budget to send a constant data stream to the cloud. This means that most sensor data will be forced to be discarded, even if it contains useful information.
Bandwidth limitations are common. This is not only about information transmission but also about power consumption. Network communication is often the most energy-consuming task in embedded systems, which means battery life also limits this functionality. Some machine learning models can be very computationally intensive, but they often use less energy than transmitting signals. This is where Edge AI comes into play. If we can run data analysis on the IoT device itself without having to upload data to obtain predictions of impending machine failures, we can use limited bandwidth to send notifications. This is much more feasible than trying to stream all data.
Of course, it is also common for devices to have no network connection at all! In this case, Edge AI enables complete application scenarios that were previously impossible.
Latency
Transmitting data takes time. Even with ample available bandwidth, the round trip from device to server can take tens or hundreds of milliseconds. In some cases, latency can be measured in minutes, hours, or even days, such as in satellite communications or message forwarding.
Some applications require faster response times. For example, controlling a vehicle remotely from a server may be impractical. Controlling a vehicle in a navigation environment requires constant feedback between steering adjustments and vehicle position. With significant latency, steering becomes a major challenge! Edge AI does not require data round trip time at all. For instance, in autonomous vehicles, the AI system runs on the onboard computer. This allows it to respond immediately to changing external conditions, such as a driver ahead suddenly slamming on the brakes.
Economics
Data connectivity can be costly. The operational costs of connected products are higher, and the infrastructure they rely on requires funding from manufacturers. The more bandwidth required, the higher the costs. This is especially true for remote devices that need to connect via satellite. By processing data on the device, Edge AI systems can reduce or avoid the costs associated with transmitting data over the network and processing it in the cloud. This can unlock many previously unattainable application scenarios.
In some applications, the only effective “connection method” is to send someone to perform some manual tasks. For example, wildlife researchers often use cameras to monitor wildlife in remote areas. These devices take pictures when motion is detected and store them on an SD card. However, the cost of uploading each photo via satellite internet is too high, so researchers must travel to their camera preset points to collect images and clear storage. Because traditional cameras are purely motion-activated, they capture many unnecessary photos—such as branches moving in the wind, passing hikers, and other photos triggered by animals that researchers are not interested in. However, some teams are now using Edge AI to identify the animals they care about, allowing them to discard other images. This means they do not have to frequently fly to remote locations to change SD cards.
Additionally, for products relying on server-side AI, the costs of maintaining server infrastructure can complicate business models. If end devices must relay information to the cloud to make decisions, users may be forced into subscription models, and device manufacturers must commit to long-term server maintenance.
Do not underestimate the impact of economics. By reducing the costs of long-term support, Edge AI can enable a multitude of previously unfeasible application scenarios.
Reliability
Device-side AI-controlled systems may be more reliable than cloud-dependent systems. Wireless connections are a vast and extremely complex web of dependencies, involving everything from link-layer communication technologies to cloud servers running applications. Many aspects of this issue are beyond the control of device manufacturers. Even if the device makes all the right decisions, it still faces reliability risks.
For some application scenarios, this is acceptable. For example, in voice-command-based smart speakers, users understand that their devices will stop recognizing commands when their internet is interrupted. But in other cases, safety is paramount. For instance, an AI-based industrial machine monitoring system must ensure it operates within safe parameters. If it stops working when the internet is interrupted, it could endanger human lives. If AI is entirely device-based, it is safer, as it can still operate in the event of connection issues.
Reliability is often a trade-off, with the required level of reliability varying by application scenario. Edge AI can be a powerful tool for enhancing product reliability. While AI is inherently complex, its complexity differs from network interconnectivity, and in many cases, its risks are easier to manage.
Privacy
In recent years, many people have chosen to sacrifice privacy in weighing convenience against privacy concerns. Theoretically, if we want products to be smarter and more user-friendly, we must give up data. Because traditional smart products rely on decisions made on remote servers, they ultimately send sensor data to the cloud.
For some application scenarios, this may not matter, such as IoT thermostats reporting temperature data to remote servers. But for other scenarios, privacy is a significant issue. For example, installing internet-connected security cameras in the home may provide some reassuring security features, but the real-time video and audio sources from the most private spaces could be broadcast to the internet. Even if the camera manufacturer is entirely trustworthy, data is always at risk of being exposed through security vulnerabilities.
Edge AI offers an alternative. Security cameras can use onboard information to determine whether there are intruders when the owner is away at work, rather than transmitting real-time video and audio to remote servers. It can alert the owner appropriately. When data is processed on embedded systems and never transmitted to the cloud, user privacy is protected, and the potential for misuse is reduced.
The ability of Edge AI to ensure privacy can be applied in many scenarios. This is particularly important for applications in security, industry, childcare, education, and healthcare. Because some of these fields involve strict regulations or customer expectations regarding data security, products with optimal privacy are those that completely avoid data collection by deploying Edge AI.
In summary, this article summarizes the advantages and necessity of Edge AI based on five aspects: bandwidth, latency, economics, reliability, and privacy. In the future, with the development of chip technology, more and more devices will deploy Edge AI to support product performance.