Accuracy vs Complexity: A Small Scale Dynamic Neural Networks Case
Martynas Dumpis, Dalius Navakauskas · Baltic Journal of Modern Computing · 2024
This research provides a detailed analysis of small-scale dynamic neural network (NN) models for human activity recognition using data from smartphones.We evaluate eight dynamic NN: Finite Impulse Response (FIRNN), Infinite Impulse Response (IIRNN), Gamma Memory (GMNN), Lattice Ladder (LLNN), Time Delay (TDNN), Recurrent Neural Network (RNN), Gated Recurrent Unit (GRUNN), and Long Short-Term Memory (LSTMNN), utilizing a publicly available dataset from Kaggle.The study focuses on comparing these models in terms of higher accuracy, smallest scale, adaptivity to the task (walking vs running classification), and memory utilization.Different NN architectures and synapse configurations are evaluated by their accuracy and computational complexity.The findings reveal which NN architectures offer the best performance while being the least computationally and memory demanding.Among the models, the IIRNN achieved the highest accuracy at 99.86% in the recognition of specified activities.Additionally, the TDNN model demonstrated impressive performance with 99.27% accuracy while requiring fewer computational resources: 2 binary additions, 2 multiplications, and 2 activation functions.