CONIX Publication

AI-enabled IoT holds the key to making rich, robust, and intelligent inferences for time-critical applications from sensor data. Unfortunately, the form factor, communication bandwidth, and energy budget of IoT platforms are going down, while the workload and complexity of neural network pipelines are going up. The first-generation efforts in AI-IoT focused on the exploration, optimization and integration of simple neural networks to low-end IoT devices without platform-awareness. In this poster, we present THIN-BAYES, a platform-in-the-loop optimization framework for training and deploying resource efficient models as small as 6 kB directly on low-end IoT platforms. THIN-Bayes adopts a platform-in-the-loop approach to get hardware metrics without relying on existing proxies and uses a gradient-free, black-box and parallelizable Bayesian optimization search strategy. The framework allows joint optimization of neural and symbolic components, unaffected by loss contour discontinuities and have fast convergence time. Models generated by THIN-Bayes are 31-740× smaller, yet approach or exceed the resolution of state-of-the-art regression or classification models.