Publication Type

PhD Dissertation

Version

publishedVersion

Publication Date

8-2026

Abstract

Embedded and wearable devices increasingly rely on deep neural networks to support intelligent sensing applications directly on the device. However, building practical on-device intelligence remains challenging throughout the deployment lifecycle. Before deployment, models must remain accurate despite severe memory, computation, and energy budgets, and a single device often needs to support multiple sensing applications without deploying a separate model for each application. After deployment, the device continues to operate in changing real-world environments, where shifts in data distribution can degrade the model’s performance over time, so the model must be able to adapt to such shifts efficiently.

The first challenge concerns accurate inference when the model design is severely constrained. Existing approaches mainly compress a large model into a smaller one, often trading off accuracy for size. To address this challenge, this dissertation rethinks the problem from the inverse perspective: rather than compressing a single model, DiTMoS composes multiple weak but complementary models and routes each input to the most suitable one through lightweight selection. This design is built on the insight that a combination of weak models can exhibit high diversity, and that their union can substantially raise the upper bound of overall accuracy.

The second challenge concerns generalizable multi-task wearable intelligence. While recent wearable foundation models (FMs) offer a promising paradigm for diverse sensing tasks through a shared encoder, their practical deployment on resource-constrained systems remains unexplored. To address this challenge, we propose LifeFM, which formulates wearable FM realization as a three-stage lifecycle, comprising pre-deployment compression, initial task specialization, and continual evolution, in which every stage must preserve the encoder’s generalizability for future tasks.

The third challenge concerns deciding when a deployed model should adapt under domain shift. Existing test-time adaptation (TTA) methods typically adapt continuously on every incoming batch, which can be unnecessarily costly when the environment is stable. To address this challenge, OD-TTA introduces an on-demand adaptation paradigm that detects domain shift at runtime and triggers adaptation only when warranted, substantially reducing unnecessary computation in stable environments.

The fourth challenge concerns making the adaptation mechanism lightweight and architecture-agnostic. Most TTA methods rely on backpropagation, which is memory- and computation-intensive, while backpropagation-free alternatives tend to be architecture-specific or incur high latency. To address this challenge, PEA revisits domain shift from an embedding perspective, showing that it induces three structural changes in the embedding space: a translation (mean shift), a scaling (variance shift), and a rotation (covariance shift). Building on this insight, PEA corrects these embedding distortions using only two forward passes, without backpropagation, and applies uniformly to both convolutional and transformer-based architectures.

In summary, these methods provide a coherent view of on-device intelligence as a lifecycle. By addressing these four challenges across embedded and wearable devices, this dissertation advances practical methods for making on-device intelligence more deployable, transferable, adaptive, and efficient under real-world resource constraints.

Degree Awarded

PhD in Computer Science

Discipline

Artificial Intelligence and Robotics | Computer Engineering

Supervisor(s)

ZHENG, Baihua

First Page

1

Last Page

218

Publisher

Singapore Management University

City or Country

Singapore

Copyright Owner and License

Author

Download

Share

COinS