How “small machine learning” can have a big impact

How “small machine learning” can have a big impact

This article has been reviewed in accordance with Science

Fact check

trusted source

Written by researcher(s)


Map of the TinyML academic network. More than 50 universities are part of the network as of February 2024. Credit: Marcelo Rovai, CC BY-SA

× Close

Map of the TinyML academic network. More than 50 universities are part of the network as of February 2024. Credit: Marcelo Rovai, CC BY-SA

The artificial intelligence (AI) application landscape has traditionally been dominated by the use of resource-intensive and centralized servers in industrialized nations. However, recent years have seen the emergence of small, energy-efficient devices for AI applications, a concept known as tiny machine learning (TinyML).

We’re most familiar with consumer-facing apps like Siri, Alexa, and Google Assistant, but the limited cost and small size of these devices allow them to be deployed in the field. For example, this technology has been used to detect mosquito wingbeats and thus help prevent the spread of malaria. He has also been part of the development of low-energy animal collars to support conservation efforts.

Small size, big effect

Small and low-cost, TinyML devices operate within constraints reminiscent of the dawn of the personal computer era: memory was measured in kilobytes, and devices could be obtained for just $1. This is possible because TinyML does not require a laptop or even a mobile phone. Alternatively, it can be run on simple microcontrollers that operate standard electronic components around the world. In fact, with 250 billion microcontrollers already deployed globally, TinyML-enabled devices are already widely available.

A number of development packages are available for TinyML applications. Two popular options are the Arduino and Seeed Studio, both of which come with additional sensors for audio, vision, and motion-based applications.

How it works?

Like classical machine learning, TinyML involves data collection—often from Internet of Things (IoT) devices—and cloud-based training. Let’s consider an application for outdoor object detection – for example, counting the number of cars on the street to see how heavy traffic there is. In the classical machine learning process, images must be collected using a webcam and sent to a cloud server where training takes place. Once the trained model provides an acceptable level of accuracy, the system is ready to detect cars from a new video stream. The ML model runs on the cloud, so an Internet connection is required.

However, in a TinyML system, the model is deployed on the device itself and is ready for offline object detection. The first part of the process (data collection and model training on the cloud) follows the classic machine learning model but the inference phase (object detection) runs on the device itself. This is how TinyML differs from traditional server-based architectures: it deploys pre-trained embedded models optimized for limited resources on embedded devices, enabling real-time, low-power data analysis and decision-making, all independent of cloud connectivity.

TinyML offers several advantages over traditional central server-based models:

  • Affordability: The low cost of the technology makes these devices accessible to a wide range of users including educational institutions and students in the developing world.
  • Sustainability: Modest energy consumption results in a low carbon footprint, reducing the impact on the environment.
  • Flexibility and scalability: It enables the development of applications that meet the needs of local communities rather than global agendas.
  • Internet-independent: Because everything is embedded, TinyML devices can work without an Internet connection. This is especially useful for the third of the world’s countries that still do not have access to the Internet.

TinyML applications already operate dedicated track and field sensors and provide translation in places where GPS is not available. They’re also being employed by startups like Useful Sensors, which offers privacy-preserving conversational agents, QR code scanners, and people detectors. Only by using TinyML can these smart devices run on low-cost, low-power microcontrollers.

developing countries in the global south

To help grow the use of TinyML in regions where a centralized machine learning model may face significant challenges, we have built TinyML4D, a network of academic institutions in developing countries. It already includes more than 40 countries stretching across the Global South from Colombia to Ethiopia to Malaysia.

With support from the UNESCO International Center for Theoretical Physics (ICTP) and from the Harvard John Paulson School of Engineering and Applied Sciences, the network was launched in 2021. The network aims to develop a community of educators, researchers and practitioners focused on both. Improving access to TinyML education, and developing innovative solutions to address the unique challenges facing developing countries.

To make all of this possible, we needed to develop ways to share educational resources globally. Initial efforts included distributing TinyML hardware kits to selected universities facing budget challenges. We also organized global and regional workshops and training courses (Africa, Latin America and Asia). Using a mix of in-person, online and hybrid methods, we reached over 1,000 participants in over 50 countries. The combination of free or low-cost hardware resources, along with open source course materials and workshops, has made it possible for TinyML to be taught by many members of our network in their home countries.

In addition to workshops and training activities, we launched a series of regional collaborations, outreach activities and virtual “show and speak” events to share best practices and increase the impact of our network among practitioners. Throughout, there has been a strong focus on addressing the UN Sustainable Development Goals (SDGs).

These collaborations have led to numerous peer-reviewed papers on applications of TinyML. In addition to mosquito species detection, which could lead to more efficient anti-malaria campaigns, other solutions include the responsible use of smart sensors and low-cost solutions for monitoring atrial fibrillation and sinus rhythm. They are also being used by Cornell University’s Elephant Listening Project, as well as monitoring water quality in aquaculture to help make it more sustainable, a project supported by the European Union’s Horizon 2020 programme.

He looks forward

TinyML represents a transformative approach to AI and is particularly relevant to developing countries. It provides a sustainable path towards democratizing AI technology, fostering local innovation, and addressing regional challenges.

However, the growth of TinyML devices and applications is not without challenges and potential risks. The number of apps and devices is expected to rise from the millions shipped today to 2.5 billion devices in 2030, and this could lead to an increase in e-waste due to the low-cost nature of the devices. There is also a risk of biases built into critical machine learning models, since they run independently, and there is no option for updates. Finally, there are privacy concerns due to the separate integration of devices into the environment. As this field develops, it will be necessary to address these issues responsibly, thus helping to ensure that TinyML remains a tool for positive change and sustainable development.

You may also like...

Leave a Reply