Wearable and always-on devices work better when they adapt to a specific user, key- word, or condition, but the tiny processors they run on cannot be retrained the usual way: memory and computation are scarce, and the standard training procedure is too heavy to run on the device itself. This thesis asks whether a Gaussian Mixture Model (GMM), a simple and well understood probabilistic model, can act as a lightweight layer that learns a single target on the device from only a handful of examples and then stays fixed. The GMM is compared against four established alternatives on three recognition tasks: spotting a chosen keyword in speech, recognizing a wearer from activity data, and a tabular control task. Each method learns only from examples of the target and is then judged on how reliably it accepts that target while rejecting everything else, together with how much memory and computation it needs. The GMM is consistently competitive without being the best in every case, and which method wins appears to depend more on how the input data is represented than on the method alone. Because it learns without the heavy training procedure and keeps a small, fixed memory footprint regardless of how many enrollment examples are used, the GMM is a practical option for adapting models directly on resource-constrained devices.
Wearable and always-on devices work better when they adapt to a specific user, key- word, or condition, but the tiny processors they run on cannot be retrained the usual way: memory and computation are scarce, and the standard training procedure is too heavy to run on the device itself. This thesis asks whether a Gaussian Mixture Model (GMM), a simple and well understood probabilistic model, can act as a lightweight layer that learns a single target on the device from only a handful of examples and then stays fixed. The GMM is compared against four established alternatives on three recognition tasks: spotting a chosen keyword in speech, recognizing a wearer from activity data, and a tabular control task. Each method learns only from examples of the target and is then judged on how reliably it accepts that target while rejecting everything else, together with how much memory and computation it needs. The GMM is consistently competitive without being the best in every case, and which method wins appears to depend more on how the input data is represented than on the method alone. Because it learns without the heavy training procedure and keeps a small, fixed memory footprint regardless of how many enrollment examples are used, the GMM is a practical option for adapting models directly on resource-constrained devices.
Adaptive Probabilistic Models for TinyML-based Wearable Personalization
BASAGLIA, ALBERTO
2025/2026
Abstract
Wearable and always-on devices work better when they adapt to a specific user, key- word, or condition, but the tiny processors they run on cannot be retrained the usual way: memory and computation are scarce, and the standard training procedure is too heavy to run on the device itself. This thesis asks whether a Gaussian Mixture Model (GMM), a simple and well understood probabilistic model, can act as a lightweight layer that learns a single target on the device from only a handful of examples and then stays fixed. The GMM is compared against four established alternatives on three recognition tasks: spotting a chosen keyword in speech, recognizing a wearer from activity data, and a tabular control task. Each method learns only from examples of the target and is then judged on how reliably it accepts that target while rejecting everything else, together with how much memory and computation it needs. The GMM is consistently competitive without being the best in every case, and which method wins appears to depend more on how the input data is represented than on the method alone. Because it learns without the heavy training procedure and keeps a small, fixed memory footprint regardless of how many enrollment examples are used, the GMM is a practical option for adapting models directly on resource-constrained devices.| File | Dimensione | Formato | |
|---|---|---|---|
|
alberto_basaglia_2119289.pdf
Accesso riservato
Dimensione
966.8 kB
Formato
Adobe PDF
|
966.8 kB | Adobe PDF |
The text of this website © Università degli studi di Padova. Full Text are published under a non-exclusive license. Metadata are under a CC0 License
https://hdl.handle.net/20.500.12608/119135