WiMi Hologram Cloud Inc. (NASDAQ: WIMI) (“WiMi” or the “Company”), a leading global Hologram Augmented Reality (“AR”) Technology provider, today announced that a new deep learning method applied in the field of BMI with data enhancement is developed to solve some difficulties in the current BMI technology.
First of all, one of the main difficulties in the field of BMI is the insufficient amount of data. Due to the difficulty of human EEG data acquisition and the limited amount of data, it is difficult to train accurate classifiers. And the traditional data enhancement methods also have limitations. Therefore, WiMi proposed a new method that combines deep learning and data augmentation. The method combines empirical modal decomposition and wavelet neural network, which can generate a large amount of artificial EEG data using a small amount of EEG data, thus improving the accuracy and generalization ability of the classifier.
A new deep learning method developed by WiMi combines data augmentation and empirical modal decomposition techniques for classifying motor imagery signals. The method applies empirical modal decomposition to EEG frames, mixes their intrinsic modal functions to create new artificial EEG frames, and converts all EEG data into tensors as inputs to a neural network. Also, two neural networks combining CNN and wavelet neural networks are proposed to train the weights and classify the two types of motion picture signals. Wavelet neural network is a new type of neural network that utilizes wavelets instead of convolutional layers.
This novel deep learning method combines CNN and wavelet neural networks. The convolutional neural network is a common neural network structure, commonly used in fields such as image recognition, with good feature extraction capabilities. Wavelet neural network is a new type of neural network that utilizes wavelets instead of convolutional layers, which can better extract time-frequency information. The combination of the two can better solve the data classification problem in the field of BMI.
This is implemented by applying empirical modal decomposition to the EEG frames and mixing their intrinsic modal functions to create new artificial EEG frames, and then converting all the EEG data into a tensor, which serves as a complex Morlet wavelet as an input to the neural network. CNN and wavelet neural networks are used to train the weights and classify the two types of motion picture signals.
Empirical modal decomposition: each EEG frame is decomposed into a number of intrinsic modal functions (IMFs) and a cosine term by empirical modal decomposition technique. Then, the individual IMFs are mixed together to form a new artificial EEG frame for training the neural network.
Data enhancement: data enhancement methods such as rotation, translation and scaling are utilized to generate some new EEG frames for training the neural network. This expands the dataset and improves the robustness and accuracy of the classifier.
Data processing: all EEG data are converted to tensor as input to the neural network. Based on this, complex Morlet wavelets are used for feature extraction.
Neural network training: a two neural network model combining CNN and wavelet neural network for training weights and classifying two types of motion image signals. Among them, the wavelet neural network is a new type of neural network that utilizes wavelets instead of convolutional layers.
In addition, another difficulty in the field of BMI is the noise interference and individual variability of EEG signals. This method can better utilize the time-frequency characteristics of EEG signals and combine the advantages of CNN and wavelet neural network, which improves the robustness of the classifier to noise and reduces the influence of individual variability, thus improving the accuracy of the classifier.
For more such updates, follow us on Google News Martech News