July 2023
Dimensions of AI-enabled Mobile and Embedded Devices: an Integration Design Guide
A design guide and graphical tool for choosing AI deployment approaches on resource-constrained mobile and embedded devices.
- Context
- Master's thesis in Computer Science and Engineering at the Polytechnic University of Milan.
- Role
- Thesis author
- Contribution
- Defined five design dimensions—resources, availability, accuracy, timeliness, and privacy—and developed a graphical tool to compare local, distributed, and hybrid deployment approaches. Applied the method to an OCR use case for credit-card registration in a banking app.
- Outcome
- The thesis reports a local OCR benchmark using TensorFlow Lite and concludes that local inference fits the use case's latency and privacy needs. Its dynamically quantized model pipeline measured 28.4 MB total model size and around 0.4 seconds of inference on a Redmi K20 Pro with four threads.
The thesis examines how resource constraints shape AI deployment on mobile and embedded devices. It organizes design decisions around five dimensions: resources, availability, accuracy, timeliness, and privacy.
The proposed graphical tool uses those dimensions to compare local, distributed, and hybrid deployment strategies. A credit-card OCR use case demonstrates how the framework can guide a deployment choice. For that case, the thesis evaluates Keras-OCR models converted to TensorFlow Lite and reports its benchmark measurements on a Redmi K20 Pro.