AutoML-Driven Co-Design of Ensemble Learning and FPGA Accelerators for Efficient Edge AI

Edge intelligence requires optimized machine learning algorithms and hardware accelerators to function efficiently in resource-constrained environments. Existing solutions often struggle to balance high accuracy, energy efficiency, and low latency effectively. We introduce CoDAT, a novel framework that leverages Automated Machine Learning (AutoML) to co-design tree-based ensemble models and FPGA-based hardware accelerators. By employing Bayesian Optimization and HyperBand (BOHB), CoDAT efficiently explores the co-design space through hardware-aware auto-tuning and model compression, framing the task as a multi-objective optimization problem. We validate CoDAT's effectiveness with a case study on risk behavior prediction in connected vehicle scenarios. The results demonstrate that CoDAT achieves a more than 95% reduction in hardware resource consumption and nanosecond-level inference times without compromising predictive performance. These findings highlight CoDAT's potential to significantly enhance computational efficiency and accuracy. Consequently, CoDAT advances sustainable and scalable edge AI solutions, with broad implications for various real-time, resource-sensitive applications.