Embedding Machine Learning enables integrating intelligence in recent application domains such as Internet of Things, portable healthcare systems, and wearable devices. This paper presents an assessment of approximate computing methods at algorithmic, architecture, and circuit levels and draws perspectives for further developments and applications. The main goal is to investigate how approximate computing may reduce the complexity and enable the feasibility of embedded Machine Learning (ML) systems. Though ML is a powerful paradigm for applications in the perceptual domain (i.e. vision, touch, hearing, etc.), their computational complexity is very high and consequently real time operation and ultra-low power are still very challenging objectives. On the other hand, approximate computing has emerged as an effective solution to reduce hardware complexity, time latency and to increase energy efficiency.

We present ESP4ML, an open-source system-level design flow to build and program SoC architectures for embedded applications that require the hardware acceleration of machine learning and signal processing algorithms. We realized ESP4ML by combining two established open-source projects (ESP and HLS4ML) into a new, fully-automated design flow. For the SoC integration of accelerators generated by HLS4ML, we designed a set of new parameterized interface circuits synthesizable with high-level synthesis. For accelerator configuration and management, we developed an embedded software runtime system on top of Linux. With this HW/SW layer, we addressed the challenge of dynamically shaping the data traffic on a network-on-chip to activate and support the reconfigurable pipelines of accelerators that are needed by the application workloads currently running on the SoC. We demonstrate our vertically-integrated contributions with the FPGA-based implementations of complete SoC instances booting Linux and executing computer-vision applications that process images taken from the Google Street View database.

Smart health home systems and assisted living architectures rely on severely energy-constrained sensing devices, such as wearable sensors, for the generation of data and their reliable wireless communication to a central location. However, the need for recharging the battery regularly constitutes a maintenance burden that hinders the long-term cost-effectiveness of these systems, especially for health-oriented applications that target people in need, such as the elderly or the chronically ill. These sensing systems generate raw data that is processed into knowledge by reasoning and machine learning algorithms. This paper investigates the benefits of embedded machine learning, i.e. executing this knowledge extraction on the wearable sensor, instead of communicating abundant raw data over the low power network. Focusing on a simple classification task and using an accelerometer-based wearable sensor, we demonstrate that embedded machine learning has the potential to reduce the radio and processor duty cycle by several orders of magnitude; and, thus, substantially extend the battery lifetime of resource-constrained wearable sensors.

Is magic real? Do swords glow if the enemy is nearby? In the demonic world of Zork, a simulated universe entices the player into a new form of problem solving.