Aptos - Creating ML Models That Fit Your Edge Device Like a Glove

Learn how to streamline edge AI model development through this 21-minute conference talk that addresses the common challenges of deploying machine learning models on edge devices. Discover a practical automation approach that eliminates the typical struggles with model zoos, missing operations, and latency limitations while bridging the gap between data science and firmware teams. Explore the core problems plaguing edge AI deployment, including demos that fail under real device constraints, foundation models that break after export, and lengthy feedback loops that consume months of development time. Understand how Aptos, an automation engine, transforms edge AI into a streamlined "data in, model out" process by exploring parameterized architecture recipes and neural architecture search techniques. Examine the system's methodology for training promising model candidates and deploying them to a comprehensive hardware farm equipped with evaluation kits. Analyze how every candidate returns concrete performance metrics including latency, per-layer timing, memory usage, on-device accuracy, and power consumption, enabling data-driven tradeoff decisions rather than speculation. Investigate the learning layer that accelerates the development process, where meta models predict runtime performance, memory requirements, and stable hyperparameter ranges before committing computational resources. Master techniques for reducing wasted time on unsuccessful approaches while converging on models that meet specific KPIs, whether targeting sub-5ms inference on i.MX 8 Plus processors, optimizing battery life, or handling non-square inputs matching camera feeds. Discover how research-backed optimization techniques including pruning, quantization, and distillation are integrated into the workflow, allowing teams to benefit from cutting-edge methods without extensive research. Learn about the system's flexibility for chip migration and NPU evaluation through dropdown hardware swaps that trigger fresh searches optimized for new hardware platforms, minimizing vendor lock-in while maintaining deployment options. Understand how this approach compresses development timelines from traditional 12-18 month projects with large teams to producing strong, deployable candidates within one to two weeks, revolutionizing the edge AI development process.