A new systematic analysis of domain generalization (DG) techniques for smartphone-based human activity recognition (HAR) reveals that individual components offer limited gains. The study, which involved over 410,000 experiments, found that while alternative training objectives rarely consistently outperform empirical risk minimization (ERM), architectural modifications like Dynamic Domain Generalization show standalone improvements. However, combining multiple DG components often yields superior results, though these gains are dependent on the specific model and data shift scenarios. The research also highlights a significant gap between current performance and potential gains, indicating a need for better model selection strategies. AI
IMPACT Highlights the need for joint design of domain generalization components and robust model-selection strategies for improved performance in real-world human activity recognition tasks.
RANK_REASON The cluster is a research paper published on arXiv detailing a systematic analysis of components and interactions for domain generalization in human activity recognition. [lever_c_demoted from research: ic=1 ai=1.0]
- arXiv
- Dynamic Domain Generalization
- Empirical Risk Minimization
- Hugging Face
- Human Activity Recognition
- Otávio Oliveira Napoli
AI-generated summary · Google Gemini · from 1 sources. How we write summaries →
