Reviewing progresses on In-Sensor AI Computing

Tiny Machine learning (Tiny ML) and Edge Artificial Intelligence (EdgeAI) are becoming essential technologies for modern embedded signal processing deployed at the edge and in resource constrained devices. Cloud-based machine learning workloads require extensive hyper-parameter tuning, along with increasingly higher computational power and storage capacity. In contrast, TinyML and EdgeAI aim to minimize these demands, enabling low-power, energy and memory-efficient solutions tailored for a huge variety of embedded edge applications. Among these devices, the ones with the lowest amount of computational and memory assets and featuring the most challenges are undoubtedly the sensors. Concurrently, these ultra tiny packaged devices are evolving from mere measurement or metro-logical chips into sophisticated meta information generators. They are becoming smaller, more computing powerful, and equipped with programmable digital signal processors (DSP) or reconfigurable hardware capable of executing Tiny ML inferences as well as learning workloads while being integrated into the same tiny package where the sensing elements are embodied. This sensor’s evolution is shifting embedded systems from the paradigm which relies on more powerful, parameter-heavy, and computationally expensive ML models to leveraging smaller and smaller, energy-efficient, and computationally lightweight solutions that aim to keep high the accuracy. Therefore, this review paper is purposely focused on sensors, inertial, environmental, time of flight (ToF), which literature proposed to integrate, in the same package, both the sensor’s element and the Tiny ML processing, including inference and, the very challenging, learning workloads. This paper refers to this sensor family as In-Sensor-AI-Computing (ISAIC). This emerging field is providing significant hardware and software progress, particularly in applications such as autonomous sensor calibration, on-device anomaly detection, classification, and human activity recognition. These advancements span multiple industry domains such as healthcare, environmental monitoring, and industrial automation. Through ISAIC, combined real-time data acquisition and processing can be achieved as soon as data are acquired, within the same package, thus reducing latency and minimizing the need for extensive data transmission to cloud systems. The paper’s findings suggest that the future of ultra tiny ML systems is heading toward creating intelligent, efficient, and cost-effective integrated sensor solutions that are seamlessly made available into everyday devices, thereby broadening the scope and impact of AI technologies at the edge.