NearCount for Model Compression on Edge Computing-Based Smart Lighting with Product-of-Sum Function

Edge computing is an emerging concept in the Internet of Things (IoT), where some processing, such as artificial intelligence (AI), is carried out on the end device, not in the cloud. However, running AI on end devices has challenges because IoT end devices usually have limited resources. Our research aim is to utilize NearCount sampling to reduce samples in the AI models we use in edge computing-based smart lighting (EdgeSL). We added a product-of-sum (PoS) function and named our novel method NearCount-PoS. After developing the novel method, we applied NearCount-PoS on the Make_Moons synthetic dataset (MMSD) to analyze its performance in reducing training data and its influence on k-nearest neighbor (KNN) prediction performance. Then, we implemented it and measured its performance on our smart lighting control dataset. Finally, we benchmarked the NearCount-PoS method with the shuffle-split edited nearest neighbor (SSENN) method. The test results show that, on the MMSD with 900 samples, NearCount with $K=6$ provides $31 \%$ samples without reducing the accuracy of KNN in making predictions. The NearCount-PoS test results on the smart lighting dataset show that NearCount with $K=159$ provides samples of $42 \%$ with a KNN prediction accuracy reduced by 0.0001 from the accuracy of KNN predictions on the original sample with a value of 0.9998. Benchmarks with SSENN show that NearCount-PoS provides higher KNN prediction accuracy at sample numbers above 9,000.