Your smartphone already tracks steps and sleep. But accurately recognizing a wider range of human activities (walking, sitting, climbing stairs) in real time, without draining resources, remains a challenge.

Dolby researchers tackled this with DAF (Distillation, Augmentation, and Filtering), a framework designed to make AI models smarter and leaner for mobile deployment. Instead of one large AI model running on your phone, DAF trains a compact "student" model by learning from multiple large "teacher" models, each expert in specific activities. This process, called knowledge distillation, transfers intelligence without transferring size.

DAF uses three key techniques:

multi-teacher distillation to match each activity with its best teacher
contrastive learning to sharpen the student's ability to distinguish confusing activities
synthetic data augmentation using VAEs and Mix-Up to handle underrepresented or difficult classes

When tested across multiple datasets, DAF consistently outperformed single-teacher and baseline models, particularly on hard-to-classify activities, while maintaining the speed of lightweight models. When two student models are trained in parallel, a smart probability filtering method selects the better prediction per activity class — boosting accuracy without adding latency.

DAF points toward a future where on-device AI is both powerful and efficient — a critical step for health monitoring, fitness, and context-aware applications on everyday devices.

Smarter Activity Recognition on Your Smartphone, Powered by Dolby Research

{{ pageEducationalSlot.headingContent[index] }}