It is mostly used for text classification and belongs to direct push transfer learning. The definition of direct push transfer learning is that given a source domain and corresponding learning tasks, a target domain and corresponding learning tasks, direct push learning aims to improve the target prediction function in the target domain by using the same knowledge in the source domain and the target domain.

Research on Key Technologies of Attitude and Gesture Perception Computing Based on Deep Learning.

Emg signal gesture recognition based on deep learning;

Without any additional information or hand-designed feature extractor, based on high-density EMG (HD-sEMG), the spatio-temporal changes of the potential field generated by muscle activity are recorded simultaneously by using the EMG signals collected by two-dimensional array electrodes. The EMG signal in HD-sEMG describes the temporal and spatial distribution of muscle activity in the electrode coverage area, and the instantaneous value of HD-sEMG presents a relatively global measurement of the physiological process involved in muscle activity at a specific time point. Instantaneous HD-Semg can distinguish different gesture patterns, and the collected HD-sEMG can describe the spatial distribution of potential, and its corresponding heat map is EMG image. The number of pixels (resolution) in the EMG image is determined by the electrode array in its acquisition equipment, that is, the number of electrodes and the distance between them (for example, an electrode grid with rows and columns of 16 can acquire an EMG image with 8* 16 pixels.

It mainly maps the original EMG signal value from (-1, 1) to (0,255), that is, where X is the original EMG signal and I is the EMG image. An 8-layer CNN structure is constructed, and the first two convolution layers of the network are used to extract the bottom picture features of the public. The author found that the instantaneous EMG signal showed different visual characteristics in different spatial positions. In different gestures, the brightness of EMG images is stronger in the middle, lower and top band areas. It is suggested to add local connection structure in the third and fourth layers (inspired by the front-end work of face recognition), because the weights of convolution templates in different spatial positions of local connection layers are not shared, which can better extract the features in different positions on the image. According to the tag with the highest recognition ratio of gesture tags in a single window, because the above experiment is only suitable for training and testing data with large amplitude of EMG signals, high gesture recognition accuracy can be obtained, so it is necessary to adopt full-wave rectification and low-pass filtering for EMG signals (full-wave rectification and low-pass filtering are widely used EMG signal amplitude estimation methods) to obtain better EMG signals.

Emg signal gesture recognition based on depth domain adaptation;

When EMG signals of training set and test set come from different acquisition sessions. Due to the interference of electrode displacement, muscle fatigue, impedance change between electrode and skin, the EMG signal is highly correlated with the acquisition session. When the trained gesture classifier is directly applied to the new session, the accuracy is usually low. Because the distribution of EMG signals in different conversations is very different, gesture recognition based on instantaneous EMG signals in different conversations can be expressed as a multi-source domain adaptive problem accordingly.

When the calibration data is not marked, this paper uses Adaptive Batch Normalization (Adabn) to adjust the gesture classifier. Assuming that the knowledge used to distinguish different gestures is stored in the weights of each layer, AdaBN does not need to adapt the gesture tags of data, but gradually updates a small number of network parameters with the increase of unlabeled adaptation data. Given the input U, BN converts it into V, where the conversion formula of the I-th input element is:

In the training stage, the mean and variance statistics of each BN layer are calculated independently for each source domain. Because BN in the training phase calculates the statistics of each data batch independently, it only needs to ensure that the samples in each data batch come from the same session.

In the identification stage, AdaBN executes the forward propagation algorithm and updates the parameters for the given unlabeled data A.

The accuracy of this method for a single frame is 30.5%, 150ms window -39.2%, while the feature set (150ms window) and linear judgment of another algorithm are 34. 1%.

Randomly select a subset of unlabeled test set (0. 1%, 0.5%, 1%, 5%, 10%) for depth domain adaptation, and then evaluate the accuracy of gesture recognition on the whole test set. Finally, after observing about 5% of the adaptive data, the accuracy reaches the peak, with 20,000 frames of adaptive data, which is about 10 second at the sampling rate of CSL-HD EMG 2048Hz.

Moreover, the adaptive algorithm does not need to observe all kinds of gestures, and the final results are 365,438+0.3% (73.2%) and 34.6% (865,438+0.4%) from 27 kinds of gestures respectively. Another method is EMG topographic map, which is defined as EMG topographic map.

Expose and criticize

Channels and frequency bands of deep belief EEG emotion recognition

Network "

In the task of emotion recognition based on EEG signals, there are irrelevant EEG signals in multi-channel EEG signals, which will not only cause noise, but also reduce the system's ability to recognize emotions. A new Deep Belief Network (DBN) is proposed to detect the key EEG channels and frequency bands of emotion recognition.

Emotional analysis is mainly carried out from behavioral and physiological reactions, because EEG has higher accuracy and objective evaluation than facial gestures. In this paper, the EEG signals of 62-channel electrode caps were recorded by ESI neural scanning system at the sampling rate of 1000Hz. Each experiment has 15 tests, and each test includes 15s prompt, 45s test and feedback, and 5s rest. The cover paper1* * evaluated 30 experiments.

First, the original EEG data is down-sampled to 200Hz, and then the noise and artifacts are filtered by a band-pass filter from 0.3Hz to 50Hz, and then the differential entropy feature proposed before [1][2] is adopted. For EEG signals with a fixed length, the differential entropy is equivalent to the logarithmic energy spectrum of a certain frequency band. It has been proved that differential entropy has the ability to identify EEG patterns between low-frequency and high-frequency energy, so differential entropy features are calculated in five frequency bands (δ: 1-3Hz, θ: 4-7Hz, α: 8- 13Hz, β: 14-30Hz, γ: 3/kloc.

Taking the denoising characteristics of 62 channels in 5 frequency bands as input, the accuracy of DBN is 86.08% and the standard deviation is 8.34%. In this paper, the key channels and frequency bands are tested by analyzing the weight distribution of training DBN. Weight is very important for identifying emotional models, because the weights of neurons that contribute a lot to learning tasks will increase, while the weights of unrelated neurons are often randomly distributed. Figure 1 shows the weights of the first layer of neural network after training.

From Figure 2, we can see that the lateral temporal region and frontal lobe region are more easily activated than other brain regions in β and γ frequency bands. Therefore, it can be concluded that the lateral temporal and prefrontal channels are the key channels, and β and γ are the key frequency bands for identifying positive, neutral and negative emotions.

As shown in Figure 3, according to the characteristics of the weight distribution of brain regions, four different electrode placement profiles were designed, including four channels, six channels, nine channels and 12 channels, among which the best average precision and standard deviation of four channels were 82.88%/ 10.92%, while the best average precision and standard deviation of all 62 channels were 83.99%/. This shows that the four opposite electrode profiles FT7, T7, FT8 and T8 are electrodes for distinguishing emotional characteristics.

[1] Duan Renning, Zhu Jianyu, Lu Baolin. Emotional classification based on differential entropy features of EEG signals [C]// International Conference on Neuroengineering. IEEE，20 13:8 1-84。

Zheng Weilin, Zhu Jianyu, Peng Yan, et al. Emotion classification of EEG based on deep belief network [C]// IEEE International Conference on Multimedia and Expo. IEEE, 20 14: 1-6.

Electroencephalogram paper (brain decoding: behavior, emotion);

Real-time naive learning of neural correlation in ECoG electrophysiology

Cortical electrophysiology related to real-time naive learning of nerves

Address:/~ blu/papers/2015/9.pdf.

Emotional Recognition of EEG Based on Principal Component Covariant Shift Adaptive Deep Learning Network

Emotional Recognition of EEG Based on Principal Component Co-shift Adaptive Deep Learning Network

Address: euro-09.pdf.

Prediction of epileptic seizures in intracranial and extracranial EEG based on recurrent neural network

Prediction of epileptic seizures in intracranial and extracranial EEG based on recurrent neural network

Time-lapse detection based on EEG with high time resolution

High time resolution detection based on EEG signal

Address:

/profile/Richard _ Jones 2 1/publication/3039266 _ EEG-Based _ Lapse _ Detection _ With _ High _ Temporal _ Resolution/links/5457 ab 030 cf 2 bccc 49 1 1 12ed