Machine learning at the extreme edge has enabled a plethora of intelligent, time-critical, and remote applications. However, deploying interpretable artificial intelligence systems that can perform high-level symbolic reasoning and satisfy the underlying system rules and physics within the tight platform resource constraints is challenging. In this paper, we introduce TinyNS, the first platform-aware neurosymbolic architecture search framework for joint optimization of symbolic and neural operators. TinyNS provides recipes and parsers to automatically write microcontroller code for five types of neurosymbolic models, combining the context awareness and integrity of symbolic techniques with the robustness and performance of machine learning models. TinyNS uses a fast, gradient-free, black-box Bayesian optimizer over discontinuous, conditional, numeric, and categorical search spaces to find the best synergy of symbolic code and neural networks within the hardware resource budget. To guarantee deployability, TinyNS talks to the target hardware during the optimization process. We showcase the utility of TinyNS by deploying microcontroller-class neurosymbolic models through several case studies. In all use cases, TinyNS outperforms purely neural or purely symbolic approaches while guaranteeing execution on real hardware.

In wearable optical sensing applications whose target tissue is not superficial, such as deep tissue oximetry, the task of embedded system design has to strike a balance between two competing factors. On one hand, the sensing task is assisted by increasing the radiated energy into the body, which in turn, improves the signal-to-noise ratio (SNR) of the deep tissue at the sensor. On the other hand, patient safety consideration imposes a constraint on the amount of radiated energy into the body. In this paper, we study the trade-offs between the two factors by exploring the design space of the light source activation pulse. Furthermore, we propose BASS, an algorithm that leverages the activation pulse design space exploration, which further optimizes deep tissue SNR via spectral averaging, while ensuring the radiated energy into the body meets a safe upper bound. The effectiveness of the proposed technique is demonstrated via analytical derivations, simulations, and measurements in both pregnant sheep models and human subjects.

Non-invasive transabdominal fetal oximetry (TFO) has the potential to improve delivery outcomes by providing physicians with an objective metric of fetal well-being during labor. Fundamentally, the technology is based on sending light through the maternal abdomen to investigate deep fetal tissue, followed by detection and processing of the light that returns (via scattering) to the outside of the maternal abdomen. The placement of the photodetector in relation to the light source critically impacts TFO system performance, including its operational robustness in the face of fetal depth variation. However, anatomical differences between pregnant women cause the fetal depths to vary drastically, which further complicates the optical probe (optode) design optimization. In this paper, we present a methodology to solve this problem. We frame optode design space exploration as a multi-objective optimization problem, where hardware complexity (cost) and performance across a wider patient population (robustness) form competing objectives. We propose a model-based approach to characterize the Pareto-optimal points in the optode design space, through which a specific design is selected. Experimental evaluation via simulation and measurement on pregnant sheep support the efficacy of our approach.