Abstract

A BCI is a system that transforms brain impulses directly into motions for computer software, such as a word processor, or a device, such as a monitor. We describe who the target users of a BCI are and how a BCI works in this section. We also discuss the various forms of BCIs, with a particular emphasis on BCIs for those with acute disabilities.

A BCI is a system that comprises sensors for monitoring brain signals (typically in the form of ‘electrodes,’ an amplifier to enhance the feeble brain waves, and a computer that converts those signals into orders to operate software programs and devices. BCIs’ components can be made portable or accessible. BCI-controlled gadgets range from assistive equipment for paralysed persons to internet-connected devices (such as a cellphone) for healthy individuals to basic videogames or entertainment tools.

Many groups are presently working on creating BCIs for a broad array of applications and user categories. Many of the apps are just designed for short-term usage and do not require permanent installation. Therapy devices to aid in the recovery of injured individuals, as well as remote controls for healthy individuals, are instances of this. Some BCIs are meant to replace a function that has been lost or degraded as a result of injuries or sickness (such as the loss of leg function due to stroke).

This paper discusses the BCI and Various techniques to BCI. A particular focus is placed on Machine learning techniques and Deep Learning techniques to BCI. Finally, we provide an overview of the Future Direction on BCI.

Keywords: BCI, techniques to BCI, machine learning, deep learning

3.1 Introduction

“The BCI is also known as Neural Control Interface (NCI). BCI is a very interesting field, active and highly demanding topic now a day. A BCI is a direct contact between the brain and a device. The National Science Foundation funded research on BCI at the University of California, Los Angeles (UCLA) in the 1970s, which was followed by a contract with the Defence Advanced Research Projects Agency (DARPA). The BCI is a form of interface that allows users to communicate with computers. Only through the channel of Brain activity, which necessitated both connections between the Brain and the Computer program. There are several approaches for BCI, including non-invasive, semi-invasive, and invasive procedures. The National Science Foundation funded research on BCI at the University of California, Los Angeles (UCLA) in the 1970s, which was followed by a contract with the Defense Advanced Research Projects Agency (DARPA) [1]. Further in this, we are going to discuss these various forms of BCI Techniques. A BCI can help a paralyzed person to convey his/her thought to the computer application, e.g., let a person is paralyzed due to some reason and he/she will want to express his/her intention to the person but he/she is not able to express their intention that what he/ she wants to express then the BCI role begins. We need to connect his/her Brain to the computer application after stabilizing the proper connection between both the Brain and computer application he/she can convey his/ her thought to the other person.

The present paper represents the various techniques of BCI, with the main focus on Machine learning Techniques to BCI, Deep learning Techniques to BCI. In Deep learning Techniques to BCI, we are going to cover the topic Convolutional Neural Network Model (CNN) and Generative Deep Learning Models [2].

3.1.1 Various Techniques to BCI

The BCI Techniques are mainly divided into three parts which are mention below:

3.1.1.1 Non-Invasive

The BCI Sensors are placed on the scalp in the Non-invasive Technique is used to determine the electrical potentials produced by the brain Electroencephalography (EEG) or Magnetoencephalography (MEG). In Non-invasive, the Electroencephalography (EEG) signal takes place electrodes kept on the scalp, on all the most external parts. The different types of representation techniques have been shown in Figure 3.1.

Because of the affordability and availability of technology, Electroencephalography (EEG) is the most widely used non-invasive approach for studying the human brain. The other Non-invasive techniques used to study the human brain are as follows:

• Magnetoencephalography (MEG)
• Electroencephalography (EEG)
• Positron Emission Tomography (PET)
• Functional Magnetic Resonance Imaging (fMRI)
• Near-Infrared Spectroscopy (fNIRS).

3.1.1.2 Semi-Invasive

The electrodes are kept on the exposed surface of the human brain in the Semi-invasive technique, the BCI. Electrocorticography (ECoG) or we can say that Electrocorticography (ECoG) uses electrodes kept on the surface of the human brain to calculate the electrical activity from the cerebral cortex. This method is used for the first time in the 1950s in the Montreal Neurological Institute [3]. This method is known as Semi-invasive however it still mandatory a craniotomy to set up the electrodes because of this reason it is necessary only when surgery is compulsory for medical reasons.

In the Semi-invasive Technique when the surgery took place the electrodes may be placed outside the epidural or under the subdural. The strip and the grid electrodes cover a huge area of the cortex approx. 4–256 electrodes. For more info see Figure 3.2.

There are also some positive sides of Electrocorticography (ECoG) which are mention below [4]:

• It has high spatial resolution and signals fidelity.
• It has a conflict with noise.
• It has low clinical risk and Robustness over a long recording period.
• It has a better-quality amplitude.

3.1.1.3 Invasive

In the Invasive Technique, the BCI the micro-electrodes are kept directly into the cortex, to measure the activity of a particular neuron. In the Invasive technique, the Intraparenchymal signal is occupied directly to the implant electrodes in the cortex. The Invasive Technique of the BCI is implanted directly into the human brain during neurosurgery. In the Invasive technique after neurosurgery, there are distributed into two units of BCI which are “Single unit BCIs” and “Multiunit BCIs”. The main distinction between these two, BCIs and multiunit BCIs is that single-unit BCIs detect signals from a single region of a brain cell, whereas multiunit BCIs detect signals from many areas of the brain cell [5]. The different layers and signal sources of the brain’s structure have been shown in Figure 3.3.

3.2 Machine Learning Techniques to BCI

Machine Learning Techniques play a huge role in the field of BCI. We all know that the human brain itself is a very complex system. The signals collected from sensors reading levels of voltage are as indirect to the real brain interfacing that we will get. But at present, it’s the best we can do without an invasive scheme that needs electrodes that go beneath the scalp. We know that every human brain structure is different to each other. Any single solution which we will find according to the interfacing problem will not work for everyone’s brain. The human brains are not having a high level of logical understanding of that level of complexity, because of which we will employ machine learning technique algorithms to give an approx. evaluation of the solution of the difficult problems with comparative speed and accuracy. As compare to the human brain machines are good at the rules to identify a pattern in difficult scenarios [6]. Scientists have developed different techniques to find patterns and features that cannot be done by human beings easily, e.g., assume that we are watching a cricket game and we paused the television on an exact shot. What if wanted to predict the shot at which we have paused the television is going to be six, four. How could we do this? The easiest way is to draw a structure between the shot and the direction where the ball is going and at which speed. A machine can work in any dimension or many dimensions at a single time whereas a human brain can’t.

In this, Classifier receives the Brain signal as Input wave and indicates label on it or we can say put it in the category with the label Move right have been represented in Figure 3.4.

There are many different types of Machine learning algorithms are available but we are going to discuss only two machine learning algorithms which are the most common algorithm for the BCI which are mention below:

3.2.1 Support Vector Machine (SVM)

In everyday terms, the Support Vector Machine (SVM) algorithms, very much like other machine learning procedures, mean to characterize a curve fit for separating two unmistakable classes of information. In a higher-dimensional space, this curve is known as a hyperplane. To track down the right hyperplane, the algorithm goes through an enhancement interaction where the quantities of misclassified occurrences (players, in the above model) are limited. This interaction is the thing that we call “preparing the algorithm” and we’ll turn out how to do it later in this instructional exercise [7].

Just keep in mind that the easiest way to find the solution to classify two different items, like the two teams according to the example is to draw a line between them as shown in Figure 3.5.

Few important points will be kept in mind well-implanting SVM algorithm which is mention below:

• The features of the SVM method should always be numerical.
• When the features in the SVM algorithm are normalized and standardized, the SVM method’s performance improves significantly.

3.2.2 Neural Networks

There’s nothing more appropriate to figure out how to tackle an issue than a brain. The neurons fire in grouping and further reason a chain response of different neurons terminating which thusly makes a course impact and results, on the whole, the things that a mind can achieve. Neural organizations are designed according to this conduct and endeavor to recreate the capacity of a human mind to figure out how to take care of an issue [8]. The input layers and output layers of the design of the brain using linear structure are shown in Figure 3.6.

The following is next to each other of the primary legitimate portrayal of neurons and the neural organization we make in this instructional exercise. It’s not difficult to see the likeness of their designs.

3.3 Deep Learning Techniques Used in BCI

Deep learning is a classification tool that is used in several daily applications, for example, speech recognition, machine vision, computer vision, and natural language processing using a BCI [1, 10].

BCI is a kind of interface that is used to translate brain signals in form of messages or commands to communicate with other devices or other brains.

Designing a BCI, on the other hand, is an extremely complicated undertaking that needs the expertise of several areas, including computer science and engineering, neurology, neural networks, and signal processing.

Mainly two phases are required to construct such complex BCI (Figure 3.7),

Because SNR is unpleasant, due to which calibration is challenging in BCI.

The main difficulty of BCI is correctly identifying human intents. The reality is that BCI’s real-world implementation is constrained by both low classification accuracy and poor generalization potential [11].

Deep learning and brain knowledge have been employed in recent years to solve such problems. Unlike traditional machine learning algorithms, deep learning can learn complex high-level features from brain signals without the need for manual feature selection, and its accuracy scales well with the size of the training set. Furthermore, Deep learning algorithms have been utilized to investigate a wide range of BCI signals (e.g., ERP, fMRI, spontaneous EEG).

Why DL is Used for BCI?

First, multiple biological, and environmental artifacts can quickly corrupt brain signals.

Working with electroencephalogram (EEG) presents several challenges. Since the primary goal of BCI is to recognize brain signals, the most common and efficient algorithms are discriminative deep learning models.

A brain signal that goes from neurons interacting with one another via the skull and scalp, and just slightly into the EEG sensor, is difficult to explain. Because EEG data is notoriously noisy, it is difficult to extract a coherent signal for a particular purpose [1, 12].

As a result, extracting meaningful information from distorted brain impulses and developing a reliable BCI mechanism that works in a range of settings are important. Furthermore, because electrochemical mind impulses are non-stationary, BCI has a poor SNR [13].

Although several pre-processing and feature development methods have been developed to minimize noise, they are time demanding and may result in information loss in the resulting features (e.g., feature extraction and selection in the time and frequency domains).

Third, feature engineering is heavily dependent on human domain information. Human knowledge can aid in capturing characteristics in some specific aspects, but it is inadequate in more general situations. As a result, a technique is needed to extract representative features automatically [14].

DL is one of the best ways for extracting distinguishable characteristics automatically.

Furthermore, since the bulk of recent machine learning research relies on static data, it is unable to correctly interpret dynamically evolving brain signals. In BCI applications, working with dynamical data sources necessitates the use of novel learning methods.

There are three benefits of deep learning:

1. First, it uses back-propagation to obtain identifiable insights from unstructured brain impulses, skipping the time-consuming pre-processing and function engineering steps.
2. Deep neural networks may also employ deep structures to collect simultaneously representational rising characteristics and hidden relationships.
3. Finally, DL algorithms outperform traditional classifiers such as SVM and LDA.

It stands to reason that nearly any BCI problem may be classed as a categorization problem.

3.3.1 Convolutional Neural Network Model (CNN)

CNN is an artificial intelligence neural net related to the visual cortex. To decrease classification mistakes, it may automatically know the proper features from the data by improving the weight values of each filter via front and backpropagation. The features extraction and classification using CNN has been presented in the Figure 3.8.

The auditory cortex, like the visual cortex, is arranged hierarchically. A hierarchical organization of brain regions performs various sorts of processing on sensory information as it enters the system. Former regions, often described as the “main visual cortex,” react to fundamental characteristics such as color and orientation. More sophisticated functions, such as object recognition, will be available in the future [14].

Deep learning approaches have the benefit of not requiring any pre-processing, allowing optimum configurations to be mastered automatically. In the case of CNN’s, attribute features extraction are combined into a single, dynamically optimized structure. Furthermore, fNIRS time series data from human participants were input into the CNN. Because the convolution is done in a sliding display way, the CNN function extraction technique maintains the temporal features of the time – series acquired by fNIRS.

3.3.2 Generative DL Models

Most of the time, Generative DL models are employed to generate training datasets or to augment findings. In other words, Generative DL models aid in increasing the quality and amount of training data in the brain signal area. After the data has been augmented, discriminative models can be performed to evaluate the data. This method was created to increase the robustness and efficacy of trained dl networks, particularly when the training set is limited.

In a nutshell, generative models accept input and produce a bunch of identical data. In this part, we’ll look at the two prominent generative dl methods: Variational Autoencoder (VAE) and Generative Adversarial Networks (GAN).

3.4 Future Direction

The BCI is an important medium for users and applications to communicate. To issue commands and complete the interaction, no external devices or muscle involvement is needed. BCIs were originally designed with healthcare applications in mind, resulting in the development of assistive devices. They also aided in the restoration of mobility capacity and the replacement of damaged motor functionality for visually disabled or locked-in people. The bright future expected for BCI has inspired researchers to explore how BCI can be used to better the lives of non-paralyzed people through medical applications.

The research’s focus, however, has been expanded to cover non-medical uses. The latest experiments have focused on average users, exploring the use of BCIs as a novel input interface and the creation of hands-free apps.

Communication and Control

BCI technologies bridge the gap between the human mind and the outside world, eliminating the requirement for traditional information transmission methods. They are in charge of transmitting and receiving signals from human minds, as well as decoding their silent thoughts. As a result, they can assist handicapped individuals with communicating and writing down their thoughts and feelings through some approaches, such as spelling applications, semantic categorization, or silent voice exchange.

Medical Applications

The medical sector has a wide range of applications that could benefit from brain impulses in all stages of treatment, including avoidance, identification, diagnosis, recovery, and reconstruction.

User State Monitoring

Initially, BCI technologies primarily focused on handicapped persons with mobility or speech issues. They aimed to provide a means for specific people to connect uniquely. However, BCI finally finds its way into the realm of safe people. It serves as a physiological measurement instrument that retrieves and uses data about a person’s mental, cognitive, and affective state.

Advertisement and Neuromarketing

The marketing sector has also inspired BCI researchers’ attention. The benefits of using EEG assessment for TV ads in both commercial and political fields were explained in the study. The created attention that comes with watching activity is calculated using BCI.

Entertainment and Games

Because of leisure and gaming applications, nonmedical brain-machine interfaces are now available. Several videogames are given, including one in which aeroplanes are rendered to navigate to a specific place in a virtual world, either in 2D or 3D. Many studies have been conducted on combining the functionality of existing games with brain controlling capabilities, to provide a multi-brain gaming experience.

3.5 Conclusion

The goal of BCI is just to translate brain activity into computer instruction. One of the most difficult parts of the BCI study is the non-stationarity of brain impulses. Because of the non-stationary, identifying consistent patterns in the signals is difficult for a classifier, resulting in low classification outcomes. When combined with the CNN and RL agent or other feedback methods, this might improve learning efficiency for operating a BCI and reduce training time.

BCI technology is rapidly evolving. They are developing methods for recording and distinguishing brain cell signals. BCI combined with caring AI can help disabled people communicate, monitor their environment, and regain their mobility. Allow paralyzed people to use their minds to operate prosthetic limbs. Will deliver auditory data to a deaf person’s head, enabling them to understand. It has a wide variety of non-medical uses. Despite its initial success, however, brain-machine interfacing poses various challenges.

References

1. 1. Alom, M.Z., Taha, T.M., Yakopcic, C., Westberg, S., Sidike, P., Nasrin, M.S., Hasan, M., Van Essen, B.C., Awwal, A.A.S., Asari, V.K., A state-of-the-art survey on deep learning theory and architectures. Electronics, 8, 3, 292, 2019.
2. 2. Chaudhary, U., Birbaumer, N., Ramos-Murguialday, A., Brain–computer interfaces for communication and rehabilitation. Nat. Rev. Neurol., 12, 9, 513–525, 2016.
3. 3. He, Y., Phat Luu, T., Nathan, K., Nakagome, S., Contreras-Vidal, J.L., A mobile brain-body imaging dataset recorded during treadmill walking with a brain–computer interface. Sci. Data, 5, 1, 1–10, 2018.
4. 4. Carabez, E., Sugi, M., Nambu, I., Wada, Y., Identifying single trial event-related potentials in an earphone-based auditory brain–computer interface. Appl. Sci., 7, 11, 1197, 2017.
5. 5. Müller, K.-R., Krauledat, M., Dornhege, G., Curio, G., Blankertz, B., Machine learning techniques for brain–computer interfaces. Biomed. Tech., 49, 1, 11–22, 2004.
6. 6. Gonfalonieri, A., A beginner’s guide to brain–computer interface and convolutional neural networks, Towards Data Science, November 25 2018.
7. 7. Rajesh Kumar, D. and Shanmugam, A., A hyper heuristic localization based cloned node detection technique using GSA based simulated annealing in sensor networks. Cognitive computing for big data systems over IoT, Springer, Cham, pp.307-335, 2018.
8. 8. Silvoni, S., Ramos-Murguialday, A., Cavinato, M., Volpato, C., Cisotto, G., Turolla, A., Piccione, F., Birbaumer, N., Brain–computer interface in stroke: A review of progress. Clin. EEG Neurosci., 42, 4, 245–252, 2011.
9. 9. Nijholt, A., Brain–computer interfaces and human-computer interaction, in: Brain–Computer Interfaces, pp. 3–19, Springer, London, 2010.
10. 10. Dhiviya, S., Malathy, S., Kumar, D.R., Internet of Things (IoT) elements, trends and applications. J. Comput. Theor. Nanosci., 15, 5, 1639–1643, 2018.
11. 11. Santhanam, G., Ryu, S., II, Yu, B.M., Afshar, A., Shenoy, K.V., A high-performance brain–computer interface. Nature, 442, 7099, 195–198, 2006.
12. 12. Graimann, B., Allison, B.Z., Pfurtscheller, G., Brain-computer interfaces: Revolutionizing human-computer interaction. 1st ed. Berlin Heidelber: Springer Verlag, 2010.
13. 13. Krishnasamy, L. A. L. I. T. H. A., et al. A geodesic deployment and radial shaped clustering (RSC) algorithm with statistical aggregation in sensor networks. Turk. J. Electr. Eng., 29.3, 1464–1478, 2021.
14. 14. King, C.E., Wang, P.T., Chui, L.A., Do, A.H., Nenadic, Z., Operation of a brain–computer interface walking simulator for individuals with spinal cord injury. J. Neuroeng. Rehabil., 10, 1–14, 2013.