Lightweight Human Activity Recognition with Model Distillation
One-line Summary
Compressing wearable-sensor activity recognition models through teacher-student distillation for edge devices.
Problem
Human activity recognition models often need to run on devices with limited compute and power. A large model may perform well during experimentation, but it can be too expensive for edge deployment. This project explores teacher-student knowledge distillation as a way to preserve useful accuracy while reducing computational cost.
System Architecture
The workflow is organized around:
- Wearable sensor time-series data preprocessing
- Teacher model training for stronger baseline performance
- Student model training through distillation
- Evaluation on activity recognition datasets such as PAMAP2 and MHEALTH
- Edge-oriented model selection based on inference efficiency
My Contribution
- Built the model distillation workflow for time-series activity recognition
- Worked with sensor data preprocessing and classification experiments
- Compared model behavior through the lens of edge-device deployment constraints
- Framed the project around practical lightweight inference rather than only raw model accuracy
Technical Stack
Python TensorFlow scikit-learn NumPy Pandas Time-Series ML Model Distillation
Results / Current Status
The project demonstrates a practical edge AI direction: start with a stronger teacher model, then train a smaller student model that is better suited to resource-constrained wearable or embedded devices.
What I Learned
- How teacher-student distillation can transfer behavior from a larger model to a smaller one
- Why time-series preprocessing is central to wearable sensor intelligence
- How deployment constraints change the way model quality should be evaluated