Lightweight Deep Learning on Edge Devices - Energy Efficient Approaches

Learn about efficient deep learning implementation on edge devices in this 32-minute conference talk from DSC ADRIA 2023. Explore research from the University of Ljubljana focusing on developing imprecise yet effective deep learning techniques that optimize power consumption and speed on smartphones, smartwatches, and IoT devices. Discover how computational accuracy requirements can be adjusted to achieve a 15% energy reduction without compromising model performance. Follow the journey from understanding AI on edge devices through scaling laws and approximate mobile computing to practical implementation using Mobiprox. Examine real-world applications in spoken keyword recognition and gain insights into strategies for balancing computational demands with device limitations. The presentation specifically targets mobile computing practitioners and includes a Q&A session addressing implementation challenges.