Neural Network Alternate Incremental Learning Algorithm for Intelligent Human Activity Recognition System
Elena S. Abramova, Alexey Orlov · 2023
In this paper, we present an alternate incremental learning algorithm for neural networks in the context of human activity recognition. The algorithm aims to address the challenges of adapting to new tasks while preserving knowledge of previously learned tasks. The proposed algorithm involves two main states: functioning and sleeping. In the functioning state, the neural network learns new tasks by adjusting the weights in the output layer while preserving the existing knowledge. However, as training progresses, the network may face difficulties in finding a balance between old and new tasks. To overcome this, the network transitions to the sleeping state. In the sleeping state, the network undergoes two phases. In the first phase, the network is trained using backpropagation, where random values are assigned to the weights connecting the input and hidden layers. Pseudo-examples are generated to enhance adaptation. In the second phase, the extreme learning machine is used to further train the network based on the reconstructed matrix P, which represents the network's knowledge. Experimental results show that the proposed algorithm achieves promising performance. The MSE values decrease significantly in the sleeping state, indicating improved adaptation and generalization capabilities. However, there is a slight increase in MSE when the network learns new knowledge, suggesting the need for further investigation and optimization. Thus, the alternate incremental learning algorithm demonstrates potential for improving the performance of neural networks in human activity recognition systems, providing a framework for adaptive learning and knowledge preservation. Future research should focus on fine-tuning the algorithm.