From wearable sensors to industrial monitors, the surge in internet-connected devices requires more sophisticated devices' artificial intelligence. Deploying complex algorithms on these resource-constrained platforms presents important challenges that drive innovation in areas such as model compression and specialized hardware. Researchers are now deploying a more powerful yet compact “small deep learning” (TinyDL) architecture that goes beyond the simple machine learning model known as “Tinyml.” A comprehensive study of this evolving field, titled “From small machine learning to small deep learning,” shows Sriyank Somvansi, MD Monzrulu Islam, Gaurav Chetri, Rohit Chakraborty, Mahmuda Sultana Mimi, The Sifatur Sifatur Isfedo Ahived Ahib, Rafat, Anandi Dutta and Subasish Das all belong to Texas State University. Their work details architectural innovation, hardware development, and software tools that drive the deployment of deep learning on strictly limited devices, along with reviews of applications across vision, audio, healthcare and industrial sectors.
Rising intelligence at the edge
The surge in embedded systems and the Internet of Things requires increasingly sophisticated data processing capabilities at the network edge, moving beyond simple sensor measurements to complex inferences. Tiny Machine Learning, or Tinyml, addresses this demand by enabling the deployment of deep learning models on resource-constrained devices such as microcontrollers. Featuring limited processing power, memory and energy budgets, these devices present a major challenge to traditional machine learning approaches.
The core principle behind Tinyml is model optimization. Typically, large size, computationally intensive, deep learning models require substantial adaptation for effective deployment on edge devices. Methods such as quantization that reduce the accuracy of numerical representations in models are of paramount importance. For example, converting a 32-bit floating-point number to an 8-bit integer dramatically reduces both the model size and the computational requirements, even if there is some degree of accuracy. Pruning, the systematic removal of redundant connections within neural networks, further contributes to the compression and acceleration of the model. These methods, often used in combination, aim to balance model accuracy with resource efficiency.
Hardware plays a key role in realizing TINYML's potential. While general purpose microcontrollers can run machine learning models, dedicated neural accelerators offer significant performance improvements. These specialized processors are designed to efficiently perform other operations central to matrix multiplication and deep learning. The development of these accelerators has driven the expansion of TINYML applications along with advances in low-power storage technologies. Edge computing is a paradigm of processing data close to its source, essentially linked to TINYML, reducing potential and bandwidth requirements compared to cloud-based solutions.
The software toolchain is evolving to facilitate the development and deployment of TINYML models. These toolchains include frameworks for model training, optimization, and compilation, converting high-level code into machine-executable instructions for specific edge devices. Automated Machine Learning (Automl) acquires traction and automates the complex processes of model selection and hyperparameter tuning. Compilers are extremely important in optimizing the target hardware model, leveraging specific architectural features to maximize performance and minimize energy consumption.
TINYML applications are diverse and expanding. Vision-based applications such as image recognition and object detection are common in areas such as smart cameras and autonomous systems. Audio recognition, including voice processing and keyword spotting, enables voice-controlled devices and acoustic surveillance systems. Healthcare applications such as wearable health monitors utilize TINYML for real-time data analysis and personalized health insights. Industrial monitoring, which utilizes predictive maintenance algorithms, improves efficiency and reduces downtime.
New trends push the boundaries of TINYML. A privacy-providing approach, Federated Tinyml allows you to train models of distributed data sources, such as data collected by many edge devices, without centralizing the data. Adapting pre-trained basic models typically presents major challenges for the deployment of edge devices, which are typically trained on large datasets in the cloud, but offers potential for improved performance. Domain-specific co-design where hardware and software are co-optimized for specific applications promises additional benefits of efficiency and performance. However, the deployment of machine learning models on resource-constrained devices also introduces security vulnerabilities and requires careful consideration of potential attack vectors and mitigation strategies.
