Invention Description
Personalized human-sensing systems, especially in clinical and health applications, often struggle with limited user-specific data and significant variability between individuals. Differences in physiology, behavior, and environmental conditions can cause performance drops when models trained on general datasets are applied to specific users. Traditional approaches either overfit to small personal datasets or fail to adapt adequately to individual differences. This creates a need for efficient personalization methods that remain robust across varying contexts while operating on resource-constrained devices.
Researchers at Arizona State University have developed CRoP, an adaptive pruning framework that enhances static human-sensing personalization by combining pre-trained models with dynamic parameter adjustment. CRoP selectively prunes and refines model components to capture user-specific characteristics while preserving essential general knowledge. This balance improves robustness to intra-user variability and distribution shifts, particularly in clinical and health-related settings with limited data. The approach is computationally efficient, making it suitable for on-device deployment. CRoP enables more accurate, reliable, and personalized human-sensing performance across diverse real-world scenarios.
CRoP is an innovative approach that enhances human-sensing model personalization by adapting static models to individual users while maintaining robustness across contexts.
Potential Applications
- Health monitoring and clinical diagnostics through personalized sensing technologies
- Wearable devices requiring on-device, privacy-preserving machine learning
- AI-enabled human activity recognition systems in varied environmental conditions
- Personalized fitness/wellness and activity tracking platforms
- Robust human-computer interaction platforms sensitive to user-specific behaviors
Benefits and Advantages
- Utilizes off-the-shelf pre-trained models, reducing the need for custom training and enhancing data privacy
- Adaptive pruning balances personalization and generalization for improved robustness across diverse contexts
- Demonstrates significant performance improvements across multiple datasets through empirical validation
- Efficient enough for resource-constrained environments, supporting on-device training
- Maintains user privacy by minimizing the need for continuous data collection and updates
For more information about this opportunity, please see