Zhuozheng SHI

Epilepsy is a chronic disorder that leads to transient neurological dysfunction and is clinically diagnosed primarily by electroencephalography. Several intelligent systems have been proposed to automatically detect seizures, among which deep convolutional neural networks (CNNs) have shown better performance than traditional machine-learning algorithms. Owing to artifacts and noise, the raw electroencephalogram (EEG) must be preprocessed to improve the signal-to-noise ratio prior to being fed into the CNN classifier. However, because of the spectrum overlapping of uncontrollable noise with EEG, traditional filters cause information loss in EEG; thus, the potential of classifiers cannot be fully exploited. In this study, we propose a stochastic resonance-effect-based EEG preprocessing module composed of three asymmetrical overdamped bistable systems in parallel. By setting different asymmetries for the three parallel units, the inherent noise can be transferred to the different spectral components of the EEG through the asymmetric stochastic resonance effect. In this process, the proposed preprocessing module not only avoids the loss of information of EEG but also provides a CNN with high-quality EEG of diversified frequency information to enhance its performance. By combining the proposed preprocessing module with a residual neural network, we developed an intelligent diagnostic system for predicting seizure onset. The developed system achieved an average sensitivity of 98.96% on the CHB-MIT dataset and 95.45% on the Siena dataset, with a false prediction rate of 0.048/h and 0.033/h, respectively. In addition, a comparative analysis demonstrated the superiority of the developed diagnostic system with the proposed preprocessing module over other existing methods.

Physical reservoir computing (RC), which can be implemented by various physical systems, is a low-cost neuromorphic framework with a fast learning capability. In the previous studies, an overdamped bistable system-based RC (OBRC) inspired by the FitzHugh-Nagumo neuron model has been proposed to construct an outstanding physical RC system. Benefitting from the stochastic resonance effect, the OBRC requires less power and has stronger noise robustness than many conventional physical RC systems. However, compared with conventional physical RC systems, its learning ability is not superior. To boost the performance of the OBRC, we propose an OBRC with time-delayed feedback (TOBRC). In this work, the TOBRC is implemented in a physical setting with time-multiplexing nodes design and simulated on a conventional computer. Moreover, we adopt a powerful optimization algorithm to automatically determine the optimal hyperparameters for both the OBRC and TOBRC; thus, a more precise quantitative discussion on the upper limit of the system can be made. To compare the TOBRC and OBRC, we conducted short-term memory and parity check tasks to assess the short-term memory ability and nonlinearity, which are the two core abilities of physical RC for learning. The results prove that the short-term memory ability and nonlinearity of the proposed TOBRC are 6.46 and 2.15 times higher than those of the OBRC, respectively. Moreover, the TOBRC outperforms the OBRC under different noise conditions. On the MNIST handwritten digit recognition benchmark, the TOBRC exhibited a lower error rate than the OBRC; it was comparable with that of advanced physical RC systems. Our study confirms that the TOBRC can exhibit excellent learning ability in practical problems.