Is the Overfitting in a Neural Network a Reliable Model for the Recognition of Activities of Daily Living?

Abstract

Recently human behaviour recognition based on mobile devices have become attractive to facilitate Ambient Assisted Living (AAL). Due to the diversities of data acquired from a large range of sensors that are available in the off-the-shelf mobile devices, the overfitting problems are highly concerned in the training of Artificial Neural Networks (ANN) for the recognition of Activities of Daily Living (ADL) and the detection of the associated environments. The purpose of this paper is to explore the correlation and causation of these ANN. We also aim to address problems such as how to avoid the overfitting and in what way these ANN models will affect the accuracy of the predictions from the automatic recognition of ADL and the detection of the environments. Several tests have been performed based on three different types of ANN created for our previously proposed framework. From the results, we can see that the implemented ANN models with Deep Neural Networks (DNN) implementation can provide reliable predictions, which is the type of ANN that has a higher probability for overfitting.