A quantum neural network for sequential data analysis in machine learning

Quantum neural networks have emerged as a novel machine intelligence (MI), showing increasing achievements despite its dawn of literature. Although the structural design for quantum ansatz plays a key role in the effectiveness and efficiency of neural algorithms, this problem is under-appreciated in the literature. In this work, we propose a hybrid quantum-classical neural network named EtaNET tailored for sequential data analysis. The quantified ML problem is translated to Quantum ML using quantum mechanics to understand algorithmic designs better. Then, we introduce a new primitive quantum ansatz with time encoding by a parameter-sharing mechanism to address the hierarchical structures of input data. The proposed algorithm allows us to truncate nearly half of the model complexity; thus, model trainability is improved from the algorithmic construction stage. Besides, we also introduce a cost-efficient training protocol for quantum Bayesian classifiers with lessons learned from the limitation of existing works. We will show the proofs of concept for EtaNET in different scaled datasets, ranging from large to extremely small cases with only 50 observations. We emphasize model transferability across datasets, enabling highly accurate Bayesian inference despite the scarcity of data curation. Besides, we will show that our model capacity is significantly larger than our classical counterparts. Finally, our model is translational quantum ML, as the model inference can be translated to address clinical-beneficial questions. The implementation of this work is given at https://github.com/namnguyen0510/EtaNet.