What are the main advantages and disadvantages of edge computing for AI? (2024)

Beyond the commonly cited performance benefits, a more important consideration may be security and privacy. The major force holding back AI adoption this decade will be worried about security and privacy - and we're already seeing numerous examples of individuals and corporations accidentally exposing too much data to SaaS AI providers. Beyond that, we're seeing government entities (especially in Europe) highly concerned with data locality related to these AI services. Edge computing may be one way to ensure data stays local and secure while still allowing for AI-enhanced functionality.

(edited)

While I understand that this is not top of mind for many people when thinking AI, 𝐭𝐡𝐞 𝐬𝐮𝐬𝐭𝐚𝐢𝐧𝐚𝐛𝐢𝐥𝐢𝐭𝐲 𝐞𝐟𝐟𝐞𝐜𝐭 𝐨𝐟 𝐫𝐮𝐧𝐧𝐢𝐧𝐠 𝐀𝐈 𝐨𝐧 𝐭𝐡𝐞 𝐞𝐝𝐠𝐞, 𝐰𝐡𝐞𝐧 𝐩𝐨𝐬𝐬𝐢𝐛𝐥𝐞, 𝐢𝐬 𝐬𝐢𝐠𝐧𝐢𝐟𝐢𝐜𝐚𝐧𝐭 and with the growing number of AI applications absolutely worthwhile to keep in mind and consider... 💚

Loss of data is another disadvantage of edge computing. Data that is deemed "useless" is often discarded after being processed in the edge devise. However, if the usefulness of data was incorrectly assessed or if we need the data at a later stage of the process or for a different aspect of the analysis then we get stuck as the data has been lost.

Something to think about with edge AI is model drift which can result in degraded output over time. Typically we're training models in the cloud then minifying them to work on constrained devices in the edge and then let them run there. We therefore need to have good edge MLOps to monitor model performance, or even to find ways to correct model drift in the edge itself (these types of feature are becoming available).But first step is to have a plan to monitor and address model drift.

(edited)

Depending on the way it is set up, the opposite can also be true: Edge AI can enhance the security of data: If the model runs on the device directly and is using a local vector database, data does not ever need to leave the device and be shared... and thus typically is safer. You can ofc set this up with a central model that runs locally to synchronize the edge models (share updates, propagate learnings...) too... as long as this is on-premise you can set that up very securely.Bonus: no need to worry about sharing your data with a cloud, SaaS, central AI model provider...

The Edge vs. Cloud definition seems a bit hazy to me: Edge Computing is a decentralized topology bringing computing to the data source; while this often can mean restricted devices (e.g. ECUs in cars or mobile phones), hardware is not a suitable criterium for distinguishing cloud vs. edge. An edge setup can entail rather capable devices and locations with many devices including big servers. While I see that hybrid solutions are most common so far, I do believe we will also see "edge-only" settings gaining some momentum in the future due to data privacy and security concerns, e.g. on shopfloors, office buildings, on offshore platforms.

This is a false dichotomy. Cloud is an operating model; edge is a location (proximity). A cloud can run at the edge. Full stop.

Here are three salient reasons to use edge computing for AI. Edge AI makes sense when:1. Latency and reliability are paramount.2. The huge amounts of data you need to process (e.g., 1000 4K camera feeds) make long-haul transit costs untenable.3. Local jurisdictions or enterprise requirements demand that data be kept locally.

Eventually, AI will be embedded in everything. It will be part of the fabric of our built environment, delivered as a utility to serve individuals, households and businesses. This pervasive AI is, by definition, edge. This is the endgame.

In an optimistic future, edge computing could stand as a transformative technology that significantly revolutionizes data handling and processing across industries. It promises a world of hyper-responsive, reliable, and efficient systems, bringing about groundbreaking advancements in areas such as IoT, autonomous vehicles, healthcare, and smart cities. With ongoing innovations in security and robust management tools, edge devices could become less vulnerable and easier to manage, effectively mitigating the risks associated with the technology. Advances in ML and AI could improve real-time insights and decision making and create new business opportunities.

More powerful micro processors with better prices and lower power consumption Bigger memory modules with better speed and pricesEmerge of SBC computersAll results in powerful edge devices that handle edge computing easier

(edited)

Edge computing is valuable for on-device data transformation and filtering, especially for feeding models. It can enable efficient sampling from dense data streams, thus reducing latency by sending aggregated and (hopefully) smaller payloads to more powerful compute substrates. However, this requires well-tested pre-processing code to be installed on edge devices. Also, any raw / original data would remain on the devices until cached out, limiting downstream access. A compromise involves including key metadata with filtered data to document essential characteristics without storing all of it. Note, the devices need to be flexibly updateable for bug fixes, performance enhancements, and ongoing compliance. So carefully select the edge device.