International Journal of Neural Systems

A full-fledged neural network modeling, based on a Multi-layered Nonlinear Autoregressive Exogenous Neural Network (NARX) architecture, is proposed for quasi-static and dynamic hysteresis loops, one of the most challenging topics for computational magnetism. This modeling approach overcomes drawbacks in attaining better than percent-level accuracy of classical and recent approaches for accelerator magnets, that combine hybridization of standard hysteretic models and neural network architectures. By means of an incremental procedure, different Deep Neural Network Architectures are selected, fine-tuned and tested in order to predict magnetic hysteresis in the context of electromagnets. Tests and results show that the proposed NARX architecture best fits the measured magnetic field behavior of a reference quadrupole at CERN. In particular, the proposed modeling framework leads to a percent error below 0.02% for the magnetic field prediction, thus outperforming state of the art approaches and paving a very promising way for future real time applications.

With the development of deep learning, the design of an appropriate network structure becomes fundamental. In recent years, the successful practice of Neural Architecture Search (NAS) has indicated that an automated design of the network structure can efficiently replace the design performed by human experts. Most NAS algorithms make the assumption that the overall structure of the network is linear and focus solely on accuracy to assess the performance of candidate networks. This paper introduces a novel NAS algorithm based on a multi-objective modeling of the network design problem to design accurate Convolutional Neural Networks (CNNs) with a small structure. The proposed algorithm makes use of a graph-based representation of the solutions which enables a high flexibility in the automatic design. Furthermore, the proposed algorithm includes novel ad-hoc crossover and mutation operators. We also propose a mechanism to accelerate the evaluation of the candidate solutions. Experimental results demonstrate that the proposed NAS approach can design accurate neural networks with limited size.

Motor imagery (MI) based brain–computer interfaces help patients with movement disorders to regain the ability to control external devices. Common spatial pattern (CSP) is a popular algorithm for feature extraction in decoding MI tasks. However, due to noise and nonstationarity in electroencephalography (EEG), it is not optimal to combine the corresponding features obtained from the traditional CSP algorithm. In this paper, we designed a novel CSP feature selection framework that combines the filter method and the wrapper method. We first evaluated the importance of every CSP feature by the infinite latent feature selection method. Meanwhile, we calculated Wasserstein distance between feature distributions of the same feature under different tasks. Then, we redefined the importance of every CSP feature based on two indicators mentioned above, which eliminates half of CSP features to create a new CSP feature subspace according to the new importance indicator. At last, we designed the improved binary gravitational search algorithm (IBGSA) by rebuilding its transfer function and applied IBGSA on the new CSP feature subspace to find the optimal feature set. To validate the proposed method, we conducted experiments on three public BCI datasets and performed a numerical analysis of the proposed algorithm for MI classification. The accuracies were comparable to those reported in related studies and the presented model outperformed other methods in literature on the same underlying data.

In this paper, a hybrid-domain deep learning (DL)-based neural system is proposed to decode hand movement preparation phases from electroencephalographic (EEG) recordings. The system exploits information extracted from the temporal-domain and time-frequency-domain, as part of a hybrid strategy, to discriminate the temporal windows (i.e. EEG epochs) preceding hand sub-movements (open/close) and the resting state. To this end, for each EEG epoch, the associated cortical source signals in the motor cortex and the corresponding time-frequency (TF) maps are estimated via beamforming and Continuous Wavelet Transform (CWT), respectively. Two Convolutional Neural Networks (CNNs) are designed: specifically, the first CNN is trained over a dataset of temporal (T) data (i.e. EEG sources), and is referred to as T-CNN; the second CNN is trained over a dataset of TF data (i.e. TF-maps of EEG sources), and is referred to as TF-CNN. Two sets of features denoted as T-features and TF-features, extracted from T-CNN and TF-CNN, respectively, are concatenated in a single features vector (denoted as TTF-features vector) which is used as input to a standard multi-layer perceptron for classification purposes. Experimental results show a significant performance improvement of our proposed hybrid-domain DL approach as compared to temporal-only and time-frequency-only-based benchmark approaches, achieving an average accuracy of $76.21\pm 3.77$%.

Electroencephalography (EEG) has been used to forecast seizures with varying success. There is an increasing interest to use electrocardiogram (ECG) to help with seizure forecasting. The neural and cardiovascular systems may exhibit critical slowing, which is measured by an increase in variance and autocorrelation of the system, when change from a normal state to an ictal state. To forecast seizures, the variance and autocorrelation of long-term continuous EEG and ECG data from 16 patients were used for analysis. The average period of recordings was 161.9 h, with an average of 9 electrographic seizures in an individual patient. The relationship between seizure onset times and phases of variance and autocorrelation in EEG and ECG data was investigated. The results of forecasting models using critical slowing features, seizure circadian features, and combined critical slowing and circadian features were compared using the receiver-operating characteristic curve. The results demonstrated that the best forecaster was patient-specific and the average area under the curve (AUC) of the best forecaster across patients was 0.68. In 50% of patients, circadian forecasters had the best performance. Critical slowing forecaster performed best in 19% of patients. Combined forecaster achieved the best performance in 31% of patients. The results of this study may help to advance the field of seizure forecasting and lead to the improved quality of life of people who suffer from epilepsy.

The representations of different fingers in the sensorimotor cortex are largely overlapped, which necessitate a good signal-to-noise ratio (SNR) and high spatial resolution to classify individual finger movements from one hand. Electroencephalography (EEG) recorded with disc electrodes has low SNR and poor spatial resolution. The surface Laplacian has been applied to EEG to improve the spatial resolution and selectivity of the surface electrical activity recording. Tri-polar concentric ring electrodes (TCREs) were shown to estimate the Laplacian automatically with better spatial resolution than disc electrodes. For this work, movement-related potentials (MRPs) were recorded from four TCREs and disc electrodes while 13 subjects performed real and imaginary finger movements. The MRP signals recorded with the TCREs have significantly less mutual information and coherence between neighboring locations compared to disc electrodes. The results also show that signals from TCREs generated higher accuracy compared to disc electrodes. It further shows that TCREs using temporal EEG data as features yield an average accuracy of $70.0\pm 7.7$% and $63.3\pm 5.8$% for real and imaginary finger movements, respectively, which is significantly higher than utilizing EEG spectral power changes in $\alpha$ and $\beta$ bands as features. Similarly, with the disc electrodes, it achieved highest accuracy of $46.13\pm 6.77$% and $34.03\pm 5.03$% for real and imaginary finger movements, respectively, with temporal EEG data feature.