Answers are grounded in this paper and cite their sources.
Model size, computational resources, and inference latency limit the application of Edge AI based on tabular neural models. This study evaluates knowledge distillation (KD) in MLP, TabTransformer, and TabNet for classifying the nutrient availability status of N, P, and K from multivariate Internet of Things sensor data. Experimental doses are divided into low, medium, and high classes, so predictions do not reflect laboratory measurements. Teacher, Student-KD, and Student-Scratch were evaluated using a fixed temporal split, five seeds, 95% confidence intervals, N, P, K, and EC channel ablation, and XGBoost, LightGBM, and CatBoost baselines. KD did not always improve accuracy, but resulted in parameter compression of 38.65 × in MLP, 7.55 × in TabTransformer, and 11.04 × in TabNet. The classic baseline remained competitive for Nitrogen and Phosphorus, while Student-KD TabTransformer produced higher mean accuracy and Macro-F1 for Potassium. Ablation shows that a single dominant channel does not drive predictions. PyTorch–ONNX conversion maintains 100% class agreement, and the validated configuration completes inference in 300 s without failure on an NVIDIA Jetson Orin NX. INT8 is not consistently faster than FP16. These results demonstrate that KD provides model efficiency, while field generalization requires polybag identification and laboratory validation.
No takes yet. Share an insight, caveat, or question.
Silaban et al. (2026) studied this question.
Synapse has enriched 10 related papers. Consider them for comparative context: