Today, mobile devices can process data natively. Here’s why distributed machine learning at the edge is maturing—and critical for enterprises.
When it comes up in conversation, artificial intelligence typically conjures images of high-powered servers running complex machine learning models in a data center.
These kinds of number-crunching scenarios are often used to analyze customer behavior, develop complex logistics plans or generate financial models. But number-crunching in a centralized data center is no longer an accurate picture of how artificial intelligence (AI) is deployed. More computation, including AI applications, is moving to mobile and edge devices. That’s because the growing volume and velocity of data requires new architectures to accommodate the high volumes of data being processed—at real-time speed. The default model to process this kind of data has been centralized, using cloud computing and on-premises data centers to process the massive influx of data.
But as our data-intensive processes handle even more information in fractions of a second, centralized computing architectures don’t cut it, creating delay. The round trip to transit the data takes too long, or data gets hung up in the journey. Data security is also an issue, with that round trip possibly compromising private information. As a result, edge computing is becoming a critical architecture to accommodate AI-enabled processes for predictions and recommendations, with more data being processed on the device before being sent to the cloud or a data center for storage.
Known as distributed machine learning, or federated machine learning, this approach enables data to be processed at the edge rather than in centralized clouds or data centers. This category of distributed machine learning trains models using data that resides on devices such as mobile phones. In addition to reducing the time it takes to send data to a centralized cloud or data center, distributed machine learning enables mobile phones to process and analyze data while keeping this information on a device.
In this article, we will explore how distributed machine learning has come to the edge.
There are several reasons that companies have moved AI from centralized on-premises data centers or public cloud architectures into a distributed environment that centers on mobile and edge devices. AI can provide various benefits, including preventative maintenance for equipment, insights about customer preferences and operational efficiencies. Because of these opportunities for process improvement and business innovation, AI at the edge is becoming a locus of activity for data. Let’s consider some examples of these benefits in various industries:
1. Manufacturing. Manufacturing facilities can enlist edge devices to monitor their equipment, collect data from operating machines and use that data to build predictive models that can foresee machine failure.
2. Logistics. Trucking operators can collect data on vehicle location, including data about the state of a vehicle prior to a part failure. By collecting real-time data on traffic and weather to optimize routes in real time, sensors can help identify vehicles with poor fuel efficiency and determine the root cause of the inefficiency, such as a malfunctioning part or low tire pressure.
3. Smart homes. Home appliances benefit from continuous monitoring by collecting predictive data that can anticipate failure. Continuous monitoring can also remotely diagnose problems so that repair personnel are prepared with the necessary parts and equipment when they arrive.
4. Personalized marketing. Targeted marketing has changed how businesses interact with consumers. Now add to that the ability to use location data in real time so you can send an ad or an offer to the right person at the right time, such as when they are nearby a store.
Just as business needs such as speed, responsiveness and mobility have prompted a move from centralized architectures to the edge, technical considerations also favor distributing AI between data centers and the edge. One of the advantages is that less data travels to a centralized data center. This helps with storage and transmission costs.
Less apparent, but increasingly important, is the impact on compliance. Less data stored means less data to protect in accordance with privacy and other regulations. Increasingly, data privacy law is moving toward policies that encourage enterprises to store only the data that is required to be stored.
For these business processes to take place, though, enterprise applications need the network to perform. Some AI applications must respond in real time or near real time, such as autonomous-vehicle controllers, which may collect data from nearby vehicles to enhance the quality of decision-making.
“The growing volume and velocity of data requires new architectures.”
Given all the data being generated and transmitted—which needs to be processed instantaneously—sending this data to the cloud is inefficient and delays analysis. Edge devices often benefit from analyzing data from devices in close proximity. For example, this data can enable enterprises to send targeted offers to customers based on their location to entice them into a store.
There are also opportunities for devices to collaborate. For example, a delivery driver can receive notifications via mobile device about the state of the truck. Machines in a manufacturing pipeline communicate status and adapt to changes in the process flow. In the consumer market, a centralized home-device manager can collect data from other intelligent devices in the home. Smart hubs and controllers can analyze data from multiple devices and optimize heating, cooling and power consumption.
These business and technology drivers motivate companies to use AI at the edge and in mobile devices, but AI wouldn’t be able to operate at the edge without three key enabling technologies: 5G, faster processors, and distributed machine learning platforms.
5G wireless networks will enable larger volumes of data to be transmitted. Information sent to data centers can be used to train and update machine learning models, such as a model to predict when a fleet vehicle will break down. Older wireless technologies would limit how much data could be sent to centralized machine learning processors. In addition, 5G networks can benefit from machine learning and AI, according to recent research.
Faster processors for mobile devices are also enabling AI at the edge. Various vendors have released intelligent chipsets that in some cases use power more efficiently than previous iterations. Many of these chips are smaller but are more powerful and have a dedicated component to execute the heavy computation that machine learning tasks require. These chips also further enable native AI processing on devices and distributed machine learning, rather than requiring data to be sent back and forth to the cloud.
Distributed machine learning platforms enable model training and prediction at the edge. This reduces the time to build and update models. It also reduces the computational load on centralized machine learning servers.
Business opportunities and technical challenges are driving the adoption of AI at the edge and on mobile devices using AI-accelerated chips and machine learning platforms tailored for mobile and Internet-of-Things devices.
While this area is primed for innovation, beware the challenges. For example, data subject to regulation needs to be secured. Devices that exchange data with other devices will need to authenticate those devices. This will require standardized protocols. Distributed cloud and edge platforms will be prime targets for malicious actors. Device and application developers will need to prepare for distributed denial of service and other attacks.
A combination of business drivers and technical constraints will drive AI from data centers to edge and mobile devices. Enabling hardware and software are in use today. Innovative applications will no doubt emerge from this confluence of factors, but risks, especially cybersecurity and privacy risks, should not be underestimated.
Dan Sullivan is a software architect specializing in streaming analytics, machine learning and cloud computing. Sullivan is the author of NoSQL for Mere Mortals and several LinkedIn Learning courses on databases, data science and machine learning.