A metaplastic neural network technique for human activity recognition for Alzheimer’s patients
Ahmed Zaghdoud, Olfa Jemai · 2023
Artificial Intelligence technology has made a huge leap forward. However, the current systems still fall short when faced with the catastrophic challenge of forgetting. Deep neural networks face huge limitations, including the tendency to forget learned tasks as they perform new tasks. While neuroscience research suggests that the human brain synapses can overcome this challenge through the ability to adapt based on historical information, applying similar mechanisms to deep neural networks to prevent forgetting previous data proves to be a monumental task.In this study, our focus lies on using these improved deep networks to provide assistance to people afflicted with Alzheimer’s disease, a prevalent condition among the elderly. Due to cognitive impairments, individuals with Alzheimer’s often require ongoing support to carry out daily activities necessary for their well-being. To tackle the issue of forgetfulness in deep neural networks, specifically binary neural networks, and grant them the ability of multitasking and continuous learning, we propose an effective training method. This can be achieved by treating the hidden weights as semi-modifiable variables and adapting the training methodology accordingly. The proposed remote monitoring system consists of two main components. The first component is the Human Activity and Patient Behavior Monitoring (HAR) Identification Unit, which focuses on monitoring and identifying patterns in the activities and behaviors of Alzheimer’s patients. The second component is a support unit that detects anomalies and behavioral problems and provides appropriate warnings. This system is based on augmentative learning technology. As a result of our research, we have developed auxiliary systems specifically designed for Alzheimer’s patients. Activity details are as follows. We achieved excellent results for Activity 1 (Task 1) with an accuracy of 99.51% and Activity 2 (Task 2) was also well-addressed by our system, reaching an accuracy of 96.63%. Finally, we conducted a comparative study, comparing our system’s performance complexity with previous systems using the Dem@care Dataset.