Tiny machine learning has enabled Internet of Things platforms to make intelligent inferences for time-critical and remote applications from unstructured data. However, realizing edge artificial intelligence systems that can perform long-term high-level reasoning and obey the underlying system physics, rules, and constraints within the tight platform resource budget is challenging. This dissertation explores how rich, robust, and intelligent inferences can be made on extremely resource-constrained platforms in a platform-aware and automated fashion. Firstly, we introduce a robust training pipeline that handles sampling rate variability, missing data, and misaligned data timestamps through intelligent data augmentation techniques during training time. We use a controlled jitter in window length and add artificial misalignments in data timestamps between sensors, along with masking representations of missing data. Secondly, we introduce TinyNS, a platform-aware neurosymbolic architecture search framework for the automatic co-optimization and deployment of neural operators and physics-based process models. TinyNS exploits fast, gradient-free, and black-box Bayesian optimization to automatically construct the most performant learning-enabled, physics, and context-aware edge artificial intelligence program from a search space containing neural and symbolic operators within the platform resource constraints. To guarantee deployability, TinyNS receives hardware metrics directly from the target hardware during the optimization process. Thirdly, we introduce the concept of neurosymbolic tiny machine learning, where we showcase recipes for defining the physics-aware tiny machine learning program synthesis search space from five neurosymbolic program categories. Neurosymbolic artificial intelligence combines the context awareness and integrity of symbolic techniques with the robustness and performance of machine learning models. We develop parsers to automatically write microcontroller code for neurosymbolic programs and showcase several previously unseen TinyML applications. These include onboard physics-aware neural-inertial navigation, on-device human activity recognition, on-chip fall detection, neural-Kalman filtering, and co-optimization of neural and symbolic processes. Finally, we showcase techniques to personalize and adapt tiny machine learning systems to the target domain and application. We illustrate the use of transfer learning, resource-efficient unsupervised template creation and matching, and foundation models as pathways to realize generalizable, domain-aware, and data-efficient edge artificial intelligence systems.

Inertial complex activity recognition and neural inertial navigation are challenging due to missing samples, misaligned data timestamps across sensor channels, variations in sampling rates and high model deployment costs. In this thesis, we introduce a robust training pipeline for complex activity detection that handles sampling rate variability, missing data, and misaligned data timestamps using intelligent data augmentation techniques. Specifically, we use controlled jitter in window length and add artificial misalignments in data timestamps between sensors, along with masking representations of missing data. In addition, we exploit end-to-end sequential learning, alpha-beta filters, Madgwick filters, hardware and quantization-aware Bayesian neural architecture search and a temporal convolutional neural network backbone to form the basis of scalable, real-time and sub-meter GPS-free inertial localization on wide spectrum of target resource-constrained hardware. We also provide a compact, ultra-low-power, environmentally resilient and modular sensor tag configuration that pushes the state-of-the-art in inertial odometry hardware. On average, the network found via our efficient pipeline provided 3x peak activation and 6x memory savings over the state-of-the-art neural inertial algorithms and taking at most 24 hours to train and search pareto-optimal models in the backbone search space. Moreover, we evaluate the complex activity pipeline on state-of-the-art complex activity recognition dataset, achieving test accuracies of 88% and 72% respectively for coarse and granular-activity classification while ranking 3rd in the 2020 Cooking Activity Recognition Challenge out of 78 submissions.