TinyML: machine learning for low-power applications.

In recent years, artificial intelligence (AI), machine learning (ML), and deep learning (DL) have gained significant popularity in both academic contexts and industrial applications. These technologies can achieve extraordinary performance, improving service quality compared to traditional solutions. For example, computer vision applications widely employ DL algorithms due to their exceptional results in image classification and recognition.

From Traditional Machine Learning to the TinyML Paradigm

However, such algorithms are characterized by high computational demands, which limit their use to powerful devices without memory and energy constraints. This represents a challenge when using deep learning in Internet of Things (IoT) systems, where simple devices with limited computational and energy resources are connected.

Considering the enormous progress of embedded devices following Moore’s law, it is now reasonable to integrate ML functionalities and ubiquitous intelligence into end devices with limited resources. Consequently, many researchers have focused their attention on this gap to make IoT end nodes intelligent and implement the TinyML paradigm.

The value of TinyML: This technology enables intelligence even in small, low-cost, and low-power devices, offering the typical advantages of edge computing.

On the other hand, implementing TinyML algorithms introduces several challenges, including:

• Model optimization and compression.
• Computational capacity.
• Reliability and maintenance.

Potential, advantages, and challenges of TinyML in resource-constrained systems

The main advantages are:

• Data processing near the sensor: Raw data are processed as close as possible to the source, avoiding transmission. In this way, only processed data containing useful information are sent, saving transmission energy and storage space. The latest and most advanced sensors allow data to be processed directly within the acquisition pipeline using simple yet effective ML solutions, reducing energy demand and execution time.
• Real-time execution: TinyML reduces communication latency by moving inference to the source device. This ensures real-time execution even in time-sensitive applications, where timely information exchange is crucial (such as in mobile robotics).
• Privacy preservation: By avoiding the transmission of raw data, TinyML ensures individual privacy without sending sensitive data over the Internet.
• Network independence: TinyML processes data locally, so it does not necessarily require a constantly available network. This enables its use in extreme scenarios where wireless connectivity may not be present.
• Cost efficiency: TinyML devices are inexpensive, enabling large-scale production, and are small in size, making them suitable for installation in hard-to-reach locations. These devices range from embedded GPUs and single-board computers to microcontrollers with only a few kilobytes of memory.

Emerging applications and research perspectives of TinyML

To fully exploit all these advantages, it is essential to develop ML models that are aware of the hardware used (hardware-aware). This introduces the main challenge of TinyML: implementing complex ML models, such as deep neural networks (DNNs), on end devices with limited computational capabilities.

Fortunately, there are solutions for model compression and optimization that allow complex models to be implemented on resource-constrained devices. However, these techniques introduce performance degradation; therefore, it is essential to optimize models while maintaining acceptable quality of service. For this reason, researchers are actively working in this field to find solutions capable of drastically compressing models without significantly compromising performance.

Many applications can benefit from TinyML, for example:

• Precision agriculture: Agricultural fields have limited resources and require low-impact monitoring systems. TinyML can provide valuable support in this scenario by integrating advanced monitoring for pests and weed infestations, thus optimizing pesticide use.
• Robotics: Mobile robotics is becoming increasingly autonomous thanks to advanced sensors and navigation techniques. However, small robots have limited payload capacity and energy resources, making TinyML essential for implementing autonomous navigation with advanced data processing.
• Industrial inspection: Companies are replacing inspection systems based on expensive devices with low-power, low-cost hardware. In this scenario, TinyML supports the integration of efficient monitoring systems into affordable devices with learning capabilities.