Edge AI refers to the execution of AI algorithms near or on devices such as wearables, mobiles, cameras, IoT (Internet of Things) sensors or edge servers, which the ability to analyse and process data. In real-time and without depending on a server on the internet.
In contrast, the use of Artificial Intelligence in Cloud (Cloud AI) requires sending data to a data center or Cloud platform where they are stored and processed. It usually requires a permanent, robust, and capable connection.
Edge AI is therefore closely related to Edge Computing. Both concepts refer to processing and analysing data at the edge of the network. That is, processing takes place in close proximity to or even on the very devices that generate the data. With Edge AI, response times are shortened, and efficiency and security are improved when executing Artificial Intelligence models on Edge devices.
Edge AI features
In order for Edge AI to be possible, analytical models and AI algorithms must be able to run on local servers or on devices equipped with processors with a certain computational capacity, as described in the article Edge Computing and Machine Learning, a strategic alliance.
However, Edge AI devices often have limitations in terms of computing power, memory, and storage space for data. Also, autonomy if they run on a battery, depending on the case.
This limitation may require optimizing algorithms and, depending on their purpose, using specialized hardware.
- Optimising algorithms by applying techniques such as reducing the complexity of models, quantifying parameters, or eliminating unnecessary connections in the neural network (pruning) to speed up its operation. Also, using of specific architectures for low-power devices.
- Specialised hardware that makes use of specific processors for AI applications, such as computer vision and neural network inference processors. AI-focused ASIC (Application-Specific Integrated Circuit) chips, such as the IBM AIU processor, offer high performance and low latency compared to general purpose processors or CPUs.
Security is also a critical aspect for Edge AI devices. Processing and analysing data on a local device reduces the risk of data exposure. But Edge AI devices are also vulnerable and equally exposed to cyber threats and attacks.
The use of Edge AI can also have privacy implications for users, depending on their function. Therefore, Edge AI must implement mechanisms for encryption, authentication, anomaly detection and privacy management.
Differences between Edge AI and Cloud AI
As mentioned above, the main difference between Cloud AI and Edge AI is in where the algorithms runs:
- In Cloud AI, data storage and processing happens on conventional, centralised Cloud servers, often managed by a Cloud provider.
- In Edge AI, data and algorithms are stored and run on hardware peripheral to the network; from wearables to vehicles to IoT devices to Edge servers, such as Snowball de AWS.
This difference means that each approach has its advantages and disadvantages as we will see more in a future post, but they can be briefly summarised as follows:
- Advantages: it usually has a greater capacity and computing power, which allows it to process and analyse large amounts of data. As a cloud platform, the power and capacity can be scaled up or down as needed.
- Disadvantages: The need to transfer data can cause speed, security and regulatory, data residency or privacy issues. Physical distance of hundreds or thousands of kilometres or network congestion can slow communications and increase data exposure.
- Advantages: requires less bandwidth and has almost imperceptible latency. It offers immediate response and increased data security, but also has an impact on users’ privacy. It can operate in remote areas or isolated environments without connectivity.
- Disadvantages: As mentioned above, limited computing power and memory can reduce the types and complexity of AI models they can run. The need to optimise algorithms and use specific hardware can increase development and deployment costs.
Edge AI applications
Edge AI can be applied in a variety of sectors. Some of the most obvious applications include security surveillance, voice assistants, Smart Industry or wearables. Some use cases would be:
- Security cameras with machine vision capabilities that analyse image content in real time to identify suspicious objects or people, events or behaviour.
- Voice assistants learn to recognise speech locally to detect when to activate or to respond to simple instructions.
- In Smart Industry, where sensors or vision systems monitor the quality of production or logistics processes or products, such as the demo we showed together with AWS at MWC.
- Wearables and medical devices, such as heart rate or glucose monitors, to analyse patient data in real time and alert on any anomaly.
Edge AI can also be applied to driver assistance systems (ADAS). Or in autonomous cars, to process data captured by vehicle sensors, such as cameras or accelerometers, to detect and respond to traffic conditions.
Edge AI is a suitable approach in cases where data needs to be analysed and processed and an immediate response is required. Without delay due to latency or insufficient connectivity or bandwidth, while maintaining data privacy and security with the right implementation.
Edge AI is still in its infancy with the continuous improvement of processing and storage capabilities in IoT devices and new developments in AI-specific hardware, such as AIU processors, Edge AI is still in its early evolutions.
The combination of Edge AI and next-generation connectivity, such as 5G and LPWA or NB-IoT via satellite, will accelerate innovation and has the potential to extend the reach of AI-enabled solutions to a wider range of industries and regions.
Featured photo: Luke Chesser / Unsplash. Apple Watch is an example of an Edge AI (consumer) device capable of processing data and executing AI models locally.