Today, as artificial intelligence continues to make breakthroughs, more and more areas begin to use AI to solve problems. Recently, Norman Di Palo, a student of artificial intelligence and robotics at Sapienza University of Rome, used deep neural networks to analyze the brain's brainwave signals and visualize and decode brain activity to understand the brain's activity intentions. . The ultimate goal of this research area is to develop affordable and useful prosthetic devices that use the brain to control prostheses to help amputees regain their ability to perform basic tasks with ease. The nervous system is a very complex structure. In your entire body, more than 100,000 kilometers of nerves communicate each part of it with your spinal cord and brain. This "grid" transmission controls the electrical impulse of each action. Each of these commands starts with your brain, which is a more magical neuron structure that communicates with electrical activation signals. Understanding and interpreting the brain's electronic patterns is one of the greatest explorations by neurologists and neurobiologists, but it turns out to be a challenging task. Electroencephalography (EEG) is a non-invasive method of recording brain activity. This is a technique that can record brain voltage fluctuations using electrodes placed on the patient's scalp. Usually, about 30 of these electrodes are placed around the scalp so that the global activity of brain waves can be recorded. In any case, the relationship between brain activity and EEG signals is complex and people know little about it without specific laboratory tests. A big challenge is learning how to "decode" (in a sense) these EEG scans, which can use a non-invasive brain-computer interface (BCI) to control robotic prostheses. And other equipment. Example of brain waves recorded with EEG As a powerful data-driven discipline, the latest breakthrough in deep learning in related pattern recognition tasks has created a new way of using neural networks to analyze these electrical signals. In this article, we will see an introduction to this topic: We will read an EEG data provided by the Kaggle contest to detect which EEG modes correspond to specific arms and gestures (eg, grabbing or lifting From the object). Then we will design a neural network that performs such classification after preprocessing the data in different ways. I will also show some data visualizations of brain activity to give a rough idea of ​​the data we are using. The ultimate goal of this research area is to develop affordable and useful prosthetic devices that can be used to control prostheses with the brain to help amputees regain their ability to perform basic tasks with ease. A similar technique can also be applied to the reading muscle electronic activation to decode what a person is trying to perform by analyzing the activated muscles. Data introduction If you have a Kaggle account, you can download data for free. As you will see, the data is made up of several .csv files. These files are: EEG data is used as an input to the model and 32 electrodes placed on the patient's scalp are used for recording. Data records are recorded at 500 Hz. In six possible situations, the human-testers tried to implement a frame-wise label. This data is collected by recording EEGs of different human testers performing simple actions (such as grabbing and lifting objects). Therefore, datasets are divided into different scenarios and have different themes. As we will see later, in the prediction of accuracy, the brain waves may be quite personal, because the model can predict exactly the same person's intention in invisible scenes, but if the training is not sufficient, they may also be very It's hard to do the same thing with new testers. Therefore, our goal is to create a neural network that takes the brain wave reading as input and outputs the probability distribution of the six possible actions the tester is trying to perform. Since "no action" is not a possible class, we can add it as a class or set all possible outputs to a value between 0 and 1, and use a threshold to determine if the action is detected. If every action is below the threshold, we think it is no action. Electrode position For better explanation, I made an animation data visualization of these electrode activities. Since the sampling frequency is very high (500 Hz), I use a simple 3-step low pass filter to smooth the data and create the first 100 frames of animation, which is about 1/5 second. Activates 32 electrodes in the first 1/5 second We can also visualize the time data as a 2D heat map, where the vertical axis is time (starting from the top and continuing downwards) and the horizontal axis represents 32 electrodes. EEG time domain heat map (time starts at the top and continues downwards) This is also very useful because we will see that it will enable us to use spatio-temporal convolutions. Data preprocessing These raw data should be preprocessed to make the learning phase more mature. For example, compared to the relatively low rate of change of the actions that have been performed, the very high sampling frequency of the EEG may cause many problems: the data changes very fast, but the action actually remains the same, so the fluctuations can be almost artificially noise . In addition, the time model receives a lot of rapidly changing data, and the classification output never changes. The first possible step is to filter the data with a low-pass filter. Even a simple running average can work: in this way, we reduce the high-frequency changes of the data while retaining the more useful low-frequency structure because the frequency of the action we are about to classify is very low (up to 1Hz). After that, we can sub-sample the data, that is, only one data point can be reserved every 10,100 times, and so on. This also helps to reduce the time dimension and reduce the relevance of the data, so that in a sense, making time more sparse. We can use many other preprocessing techniques, but for the sake of simplicity, we can stop here and start designing our neural network. Neural Network Design and Experiment When dealing with temporal data, the first structure we think of is the recurrent neural network. These networks have dynamic structures that allow them to encode temporal data, so they also calculate their output based on past inputs. I designed a long-term short-term memory network (LSTM) based on Keras and used a timing structure to input training data into the LSTM network. The result is good, but in this particular case, I am more interested in demonstrating how a convolutional neural network, often used for images, handles temporal data very well. As described earlier, in a sense, we are actually dealing with spatio-temporal data: the upper ordinate represents the time evolution of the heat map, and the abscissa represents the various electrodes, adjacent electrodes Almost always close in the physical space on the human scalp. This means that we can actually extract useful features by convolution: the 2D kernel will encode the pattern in time and space. Imagine a 3×3 convolution kernel: On the matrix described by the heat map, it can extract feature values ​​by weighted summation over 3 different time steps (3 kernel lines), but it can also be in 3 Extract features on different electrodes (3 kernel columns). Therefore, a CNN with many cores can discover how the electrode activation changes the characteristics associated with the target motion over a limited period of time. As suggested in the Kaggle contest, we examined the AUC score in order to test our model performance. If you do not understand the meaning of AUC, I suggest that you look at the clear and intuitive explanation from here (https://datascience.stackexchange.com/questions/806/advantages-of-auc-vs-standard-accuracy). Just as you can check yourself through an online notebook, after a quick training phase, we have an AUC score of about 0.85. Many improvements can be achieved by training different neural network structures and preprocessing techniques, but the introduction of this concept proves the remarkable ability of neural networks to learn from such data. in conclusion In this article, we use EEG to introduce EEG signals, a non-invasive and relatively simple method that can record the useful signals of a user's scalp. We have seen some intuitive data visualizations and how to use neural networks to extract features such as motion intent. I believe that this area (robot prosthesis, brain-computer interface) will be further promoted thanks to the increasing depth of learning and wider variety of science and technology, as well as the ever-growing platform and competition. The impact of these technologies will be enormous. Having low-cost artificial limbs that can be naturally controlled can significantly improve the lives of millions of people. Made of TPU material imported from South Korea, it can effectively protect the screen from unnecessary wear or scratches and reduce the impact of falling.
The ultra-transparent thickness of 0.14mm ensures the sensitivity of the original touch screen. When using Apple Pencil, it can still maintain a high touch sensitivity. 99% transparency provides an ideal natural viewing experience.
If you want to know more about Imported Screen Protector For iPad products, please click the product details to view the parameters, models, pictures, prices and other information about Imported Screen Protector For iPad. Screen Protector For iPad, Imported Screen Protector, 200*300 Screen Protector, Imported Protective Film, iPad Screen Protector,Imported Hydrogel Film Shenzhen Jianjiantong Technology Co., Ltd. , https://www.tpuscreenprotector.com
The flexible material design allows the Screen Protector to adapt to the contours of any device, so it can be connected to curved displays and rounded edges. The Full-Cover Screen Protector can perfectly fit your screen and provide maximum protection.
The oleophobic coating technology protects the glass from fingerprints left by sweat and grease, and fully supports the 3D touch function on the iPad.
Whether you are a group or an individual, we will try our best to provide you with accurate and comprehensive information about the Imported Screen Protector For iPad!