Active Hackathon

Enabling Artificial Intelligence on Local devices with Edge ML

Edge Machine Learning decreases dependency on Cloud networks.
Listen to this story

With the growth of the Internet of things (IoT) the Cloud networks were overburdened, and businesses ignored critical Cloud computing problems such as security. The solution for all these problems was to run Machine Learning models on local devices “Edge ML”.

Edge ML is a technology that allows Smart Devices to analyse data locally utilising machine and deep learning algorithms, decreasing dependency on Cloud networks. This article will be focused on understanding the working and functionality of Edge Machine Learning. Following are the topics to be covered.


Sign up for your weekly dose of what's up in emerging technology.

Table of contents

  1. What is IoT?
  2. What is Edge ML?
  3. How does it work?
  4. Benefits and Drawbacks
  5. Applications of Edge ML

Let’s first understand the concept of IoT and its impact of it on Edge ML.

What is IoT?

The Internet of Things (IoT) is a physical item network which connects everything to the Internet using predefined protocols and information sensing equipment.  This equipment shares information and communicates to accomplish smart recognition, locating, tracking, monitoring, and administration.

The internet is no longer just a network of computers; it has evolved into a network of devices of all shapes and sizes, including vehicles, smartphones, home appliances, toys, cameras, medical instruments and industrial systems, animals, people, and buildings, all connected, all communicating and sharing information based on predefined protocols to achieve smart reorganisations, positioning, tracing, safe & control, and even personal real-time online monitoring, online upgrade, and propulsion.

The Internet of Things (IoT) is a concept and a paradigm that considers the pervasive presence in the environment of a variety of objects that can interact with each other and cooperate with other objects via wireless and wired connections and unique addressing schemes to create new services and achieve common goals.

GIF source

Are you looking for a complete repository of Python libraries used in data science, check out here.

What is Edge ML?

Edge Machine Learning appears as a cloud computing extension to bring cloud services closer to end-users. Edge computing provides virtual computing platforms that provide processing, storage, and networking capabilities that are often positioned at the network’s edge. 

Edge servers are devices that deliver services to end devices and can be IoT gateways, routers, and mini data centres in mobile network base stations, automobiles, and other locations. Edge devices are end devices that request services from edge servers, such as mobile phones, IoT devices, and embedded devices.

The combination of edge computing with AI presents a possible answer to the major issues of AI-based applications. This new intelligence pattern is known as edge intelligence. Edge intelligence refers to a network of linked systems and devices for data gathering, caching, processing, and analysis close to where data is collected, with the goal of improving data quality and speed while protecting data privacy and security. 

Unlike traditional cloud-based intelligence, which needs end devices to upload created or gathered data to a remote cloud, edge intelligence processes and analyses data locally, effectively protecting users’ privacy, reducing reaction time, and conserving bandwidth resources.

When to use edge computing

Since, it has become more important to handle, store, and process data effectively. It is especially important for time-sensitive firms to process data quickly and effectively in order to minimise safety hazards and speed up corporate processes. Edge computing attempts to optimise web applications and internet devices while reducing bandwidth utilisation and communication delay. 

For example, Oil and gas installations are frequently located in distant areas. Edge computing allows real-time analytics by bringing processing closer to the asset, reducing dependency on high-quality connectivity to a centralised cloud.

How does Edge ML work?

The aim behind Edge ML is for the model to exist on devices at the network’s edge. The ML algorithms are then executed locally on the device, eliminating the need for an internet connection to analyse data and produce usable results. The whole process could be divided into four major components.

Edge Caching 

Edge caching is a distributed data system that gathers and stores data created by edge devices and their surrounding surroundings, as well as data received from the Internet, to support intelligent applications for users at the edge. At the edge, data is spread.

Edge devices, such as monitoring devices and sensors, collect ambient data. Such data is kept in appropriate locations and processed and analysed by clever algorithms to give services to end-users.

Caching is based on request redundancy. In edge caching, acquired data is fed into intelligent applications, and the results are transmitted back to the location where the data is stored. The redundancy could be categorised into kinds, data redundancy and Computation redundancy.

  • Data redundancy refers to the fact that an intelligent application’s inputs may be identical or partially identical. Considering a continuous mobile vision analysis, there are a lot of comparable pixels between consecutive frames. Some edge devices with limited resources may be required to transfer captured videos to edge servers for additional processing. Edge devices simply need to upload various pixels or frames when using caching. Edge devices might reuse the findings of the repeated segment to minimise excessive processing. 
  • Computation redundancy indicates that the computing tasks demanded by intelligent applications may be the same. An edge server, for example, delivers image recognition services to edge devices. Recognition tasks from the same environment may be the same, for example, object recognition tasks from various users in the same location. Edge servers might immediately communicate previously acquired recognition results back to users. 

The cache could be stored in three locations: macro and micro Base Stations, as well as edge devices. There are two types of material known as popular files and intelligent models. Macro Base Stations are typically used as edge servers in edge intelligence, providing intelligent services with stored data.

The cache is delivered, in two ways one is delivered from a single Base Station, and the other is delivered from multiple Base Stations based on the cooperation amongst them. Furthermore, the ideal content to cache is carefully explored, i.e., data redundancy based content and computation redundancy based content. When compared to macro and micro Base Stations, edge devices often have fewer resources and great mobility. Due to the limited storage capacity of macro BSs, micro BSs, and edge devices, the content replacement must be addressed.

Analytics India Magazine

Edge Training

Edge training is a distributed learning process that learns the best values for all weights and biases, as well as the hidden patterns, using the training set stored at the edge. Edge training, as opposed to traditional centralised training methods on powerful servers or computing clusters, typically takes place on edge servers or edge devices, which are typically not as powerful as centralised servers or computing clusters.

The device is trained in two ways: individually and collaboratively. Solo training is done on a single device without the aid of others, whereas collaborative training involves numerous devices working together to train a shared algorithm. Because solo training necessitates more hardware, which is sometimes unavailable, the majority of extant material focuses on collaborative training designs.

Edge training is substantially slower than centralised training paradigms, in which strong CPUs and GPUs could ensure a decent outcome with a minimum training period. Some researchers are interested in accelerating edge training. Work on training acceleration is separated into two groups, according to training architecture: acceleration for solo training and acceleration for collaborative training.

Solo training is a closed system in which the optimal parameters or patterns are obtained by iterative computing on single devices. Collaborative training, on the other hand, is based on the collaboration of several devices, which need periodic contact for updating. The frequency of updates and the cost of updates are two elements that influence communication efficiency and training success. 

Researchers in this field are primarily concerned with how to maintain model/algorithm performance with decreased update frequency and expense. Furthermore, the open nature of collaborative training makes it vulnerable to malevolent users. There is also some literature on privacy and security concerns.

Analytics India Magazine

Edge Inference

Edge inference is the stage in which a learned algorithm is employed in a forward pass to calculate the output on edge devices and servers. Most present AI models are intended to be deployed on devices with strong CPUs and GPUs; however, this is not feasible at an edge setting.

The models are made appropriate for the edge environment by either inventing new algorithms that have reduced hardware needs and are naturally suited to edge settings or compressing existing models to eliminate superfluous operations during inference.

In the case of building a new method, there are two approaches: allowing computers to construct optimum models, i.e., architecture search, and human-invented architectures using depth-wise separable convolution and group convolution.

The current models are compressed for model compression to create thinner and smaller models that are more computation and energy efficient with low or no loss of accuracy. Model compression techniques include low-rank approximation, knowledge distillation, compact layer design, network pruning, and parameter quantisation.

Analytics India Magazine

Edge offloading 

Edge offloading is a distributed computing paradigm that delivers computational functions such as caching, training, and inference at the edge. If a single edge device lacks the capacity to support a certain edge intelligence application, application responsibilities may be offloaded to edge servers or other edge devices. The edge offloading layer transparently delivers compute services to the other three edge intelligence components. Offloading strategy is critical in edge offloading, as it should make maximum use of the available resources in the edge environment. 

Cloud servers, edge servers, and edge devices house available computer resources. Device-to-cloud (D2C) offloading, device-to-edge server (D2E) offloading, device-to-device (D2D) offloading, and hybrid offloading are the most often utilised techniques.

  • The D2C offloading technique chooses to leave pre-processing operations on edge devices and offload the remainder of the jobs to a cloud server, which might cut the quantity of uploaded data and latency dramatically.
  • The D2E offloading approach employs a similar procedure, which may reduce latency and reliance on the cellular network even more.
  • Focus on smart home situations in the D2D offloading method, where IoT gadgets, wearables, and smartphones collaborate to conduct training/inference tasks.
  • Hybrid offloading solutions provide the most adaptability, making the most of all available resources.

Applications of Edge ML

Edge computing is being used in a variety of sectors. It gathers, processes, filters, and analyses data locally or at the network edge. It is used in the following areas.


Edge computing can aid with data access by utilising machine learning and automation. It aids in the identification of problematic data that requires quick attention by physicians in order to improve patient care and eliminate health occurrences.

Patients’ chronic diseases can be monitored with health monitors and other wearable healthcare gadgets. It has the potential to save lives by instantaneously informing caregivers when assistance is necessary. Furthermore, surgical robots must be able to immediately interpret data in order to help safely, promptly, and precisely. If these gadgets rely on sending data to the cloud before making judgments, the consequences might be disastrous.


Edge computing is mostly used in the construction sector for worker safety, collecting and analysing data from safety equipment, cameras, sensors, and so on. It provides organisations with an overview of workplace safety conditions and guarantees that personnel follow safety regulations.


Targeted marketing and information for retail businesses are dependent on essential factors specified in on-field equipment, such as demographic information. Edge computing can assist in securing user privacy in this scenario. Rather than transmitting unprotected data to the cloud, it may encrypt the data and maintain the source.


In the manufacturing industry, edge computing is used to monitor industrial processes and use machine learning and real-time analytics to improve product quality and detect production mistakes. It also encourages the incorporation of environmental sensors into manufacturing operations.

Edge computing also offers information on the components in stock and how long they will last. It enables the producer to make more accurate and timely business choices about operations and the facility.


Edge computing is used in sensors in farming to measure nutrient density and water consumption and improve harvesting. The sensor takes data on ambient, temperature, and soil variables to do this. It examines their impacts in order to increase agricultural productivity and guarantee harvesting occurs within the most favourable climatic circumstances.

Gas and Oil 

Edge computing may be used to monitor safety in gas and oil utilities. Sensors continually measure humidity and pressure. To respond rapidly, there needs to be adequate network connectivity. The problem is that the majority of these facilities are in distant places with limited connections.

As a result, placing edge computing at or near such systems provides increased connection and continuous monitoring capabilities. Edge computing can also detect equipment issues in real-time. The sensors can monitor energy generated by all equipment, including electric vehicles, wind farm systems, and others, and use grid management to aid in cost reduction and efficient energy creation.

Autonomous Vehicles

A pedestrian crossing in front of an autonomous vehicle must be stopped immediately. It is not fair to rely on a distant server to make this judgement. Furthermore, cars that employ edge technology can interact more effectively since they can speak with each other first, rather than first sending data about accidents, weather conditions, traffic, or diversions to a remote server. Edge computing can be beneficial.

Smart sound systems

Smart sound systems may learn to comprehend voice commands locally in order to execute simple commands. Even if internet access breaks, it would be feasible to turn lights on and off or alter thermostat settings.

Benefits and Drawbacks

The following are the advantages of edge computing.

Quicker Response Times

As previously stated, deploying compute processes at or near edge devices helps minimise latency.

Consider moving files inside the same building. It takes longer to exchange files since it talks with a remote server located anywhere in the globe and then returns as a received file. The router is in charge of data transfers throughout the workplace using Edge computing, drastically minimising latency. It also saves a significant amount of bandwidth.


Edge computing saves server resources and bandwidth, which saves money. The cost increases as you deploy cloud services to support a big number of devices in businesses or homes with smart gadgets. Edge computing, on the other hand, has the potential to lower this expense by relocating the compute portion of all of these devices to the edge.

Data Protection and Privacy

Moving data across overseas servers raises privacy, security, and legal concerns. It can pose serious problems if it is hijacked and falls into the wrong hands.

Edge computing brings data closer to its source while remaining within the parameters of data rules. It allows sensitive data to be processed locally rather than moving it to the cloud or a data centre. As a result, your data stays secure within your premises.

The following are the drawbacks of edge computing.

Storage and Cost

Even while the costs of cloud storage are reduced, there is an extra cost on the local end. Much of this stems from the development of storage capacity for edge devices. Edge computing also has a cost component since existing IT network infrastructure must be replaced or improved to support edge devices and storage. Some businesses may discover that the cost of transitioning to an edge network is comparable to the cost of constructing and maintaining a traditional IT infrastructure.

Data Loss

The benefit of edge computing comes with a danger. To minimise data loss, the system must be well planned and programmed before it is deployed. Many edge computing devices, as they should, discard useless data after collection; nevertheless, if the data removed is significant, the data is lost, and the analysis in the cloud is inaccurate.

Security risk

There is a security benefit at the cloud and corporate levels, but there is also a security danger at the local level. It is pointless for a corporation to have a cloud-based supplier with superb security if its local network is vulnerable. 


Machine learning on devices minimises network congestion by allowing calculations to be done close to data sources while maintaining privacy when uploading data. With this article, we have understood the working and applications of Edge Machine Learning.


More Great AIM Stories

Sourabh Mehta
Sourabh has worked as a full-time data scientist for an ISP organisation, experienced in analysing patterns and their implementation in product development. He has a keen interest in developing solutions for real-time problems with the help of data both in this universe and metaverse.

Our Upcoming Events

Conference, in-person (Bangalore)
Machine Learning Developers Summit (MLDS) 2023
19-20th Jan, 2023

Conference, in-person (Bangalore)
Data Engineering Summit (DES) 2023
27-28th Apr, 2023

Conference, in-person (Bangalore)
MachineCon 2023
23rd Jun, 2023

3 Ways to Join our Community

Discord Server

Stay Connected with a larger ecosystem of data science and ML Professionals

Telegram Channel

Discover special offers, top stories, upcoming events, and more.

Subscribe to our newsletter

Get the latest updates from AIM