The Size-Performance Paradox: When Tiny Models Fall Short in Edge Optimization

In the realm of edge AI, the conventional wisdom suggests that smaller models naturally translate to faster, more efficient performance on resource-constrained devices. However, my talk will challenge this assumption by unveiling the "size-performance paradox": the reality that choosing a smaller model doesn’t always yield the expected improvements in speed, energy consumption, or real-time reliability. I’ll demonstrate how hardware-aware deep learning compression and careful optimization balance model size with real-world performance, sharing practical insights and case studies for deploying AI on resource-constrained devices. Participants will gain a comprehensive understanding of how to effectively tailor AI models for edge deployment.