Count Estimation With a Low-Accuracy Machine Learning Model
Count Estimation With a Low-Accuracy Machine Learning Model
复制标题
DOI：
10.1109/jiot.2020.3038273
复制
发表时间：
2021-04
影响因子：
10.6
通讯作者：
Y. Sei;Akihiko Ohsuga
中科院分区：
文献类型：
--
作者：
Y. Sei;Akihiko Ohsuga
许多物联网(IoT)系统使用机器学习技术，例如深度神经网络。物联网系统可以使用机器学习技术预测属性，如年龄、性别、车速、人类行走速度和动物类型。虽然机器学习的功能是不可否认的，但预测精度并不总是很高。当使用机器学习模型来识别物体计数系统中的多个物体时，由于每个物体的识别误差的累积，估计计数将会有很大的误差。在这项研究中，提出了一种利用训练阶段产生的混淆矩阵进行计数估计的方法。该方法包括使用混淆矩阵迭代贝叶斯技术进行计数估计，以及减少迭代次数以减少估计误差。该方法甚至适用于低精度的机器学习模型。通过对合成数据集和真实数据集的实验，验证了该方法的有效性。实验结果表明，与基线方法相比，该方法的估计误差平均降低了64.3%。
Many Internet-of-Things (IoT) systems use machine learning techniques, such as deep neural networks. IoT systems can predict attributes, such as age, sex, car speed, human walking speed, and types of animals, using machine learning techniques. Although the functionality of machine learning is undeniable, the prediction accuracy is not always high. When a machine learning model is used to recognize several objects in an object counting system, the estimated count will have a significant error because of the accumulation of the recognition error of each object. In this study, a count estimation method that uses a confusion matrix generated in the training phase was proposed. The proposed method consists of an iterative Bayesian technique with the confusion matrix for count estimation and mitigating over-iterations technique for reducing estimated errors. The proposed method can be used even for a low-accuracy machine learning model. Experiments with synthetic and real data sets were conducted to demonstrate the functionality of the proposed method. The estimation errors of the proposed method were reduced by 64.3% in average compared to the baseline method in the experiments.