Introduction
A brain-computer interface (BCI), sometimes called a neural-control interface (NCI), mind-machine interface (MMI), direct neural interface (DNI), or brain-machine interface (BMI), is a direct communication pathway between an enhanced or wired brain and an external device. BCI’s are often directed at assisting, augmenting, or repairing human cognitive or sensory-motor functions.
Highlights of Brain-Computer Interface Research
Non-invasive BCIs that utilize electroencephalography (EEG) to record brain activity are seeing much progress. Popular EEG headset manufacturers like Emotiv and Neurosky have enabled a wide array of consumer grade BCIs for gaming, virtual reality, and other applications. Larger clinical EEG systems are utilized in therapies for paralysis, seizure disorders, and other conditions.
Invasive BCIs use brain implants to directly access neurons and neuronal pathways. These provide higher resolution signals than EEG but involve greater risks from surgery. Implantable depth electrodes and microelectrode arrays are utilized in clinical BMIs assisting with prosthetic control, in therapies for paralysis, epilepsy, and other neurological conditions.
Neural decoding algorithms are increasingly sophisticated in mapping patterns of brain activity to intended actions or stimuli. Machine learning and deep learning approaches have enabled continuous, high-dimensional decoding of brain signals recorded via EEG or invasive interfaces. Convolutional and recurrent neural networks are popular.
Brain-machine interfaces are enabling new assistive technologies for mobility, communication, environmental control, and more. Clinical trials indicate BMIs may restore lost motor function for those with paralysis or amputations. Experimental work is translating thoughts to text or allowing control of robotic arms and exoskeletons.
Hybrid BCIs integrate signals from the brain as well as other body signals like EMG. This offers expanded options for assistive device control by taking advantage of residual muscle movements or combining brain waves with limb movements.
Wireless BMI technologies remove cables obstructing natural movement. Implanted devices can transmit neural recordings via ultrasound or radio waves to external receivers. This benefits usability and broader clinical application of implanted BMIs.
Optogenetics allows for control of neurons firing patterns using light. Genetically engineered neurons in animal studies fire or do not fire in response to light, enabling new forms of neural control and stimulation. Potential long term applications in controlling prosthetics through thought.
Brain-computer interfaces can enhance normal abilities like control of advanced computers, prosthetics, vehicles or equipment. Experimental BCIs enable communication solely via thought, remote monitoring of brain activity, and neurofeedback enhancing cognitive states.
Neuroprosthetics physically interface the nervous system to restore lost functions from injury or disease. Implants process nerve signals to bypass damaged areas or stimulate muscles and senses. This is advancing toward restoring dexterous control of artificial limbs and true bionic integration.
Non-Invasive EEG Brain-Computer Interfaces
Non-invasive BCIs avoid surgery and utilize external sensors like electroencephalography (EEG) electrodes positioned on the scalp to record cortical brain waves. Key advantages include a lack of surgical risk and the ability to easily apply and remove the system as needed. This enables wider application and testing of BCIs.
EEG based BCIs commonly target sensorimotor rhythms – the Mu (8–12 Hz) and Beta (18–26 Hz) rhythms which desynchronize over relevant portions of the scalp when planning or executing movement. Through training algorithms, users can learn to voluntarily control these rhythms to operate assistive devices or computer applications.
Popular EEG headset manufacturers have unlocked a new realm of consumer-grade BCIs for normal users. Emotiv and Neurosky devices enable controlling games, apps, and virtual environments just through thought. Clinical-grade BCIs are also non-invasive but target medical applications like assistive communication for paralysis patients.
Non-invasive BCIs have benefited greatly from advanced machine learning techniques that decode complex patterns in EEG data in real-time. They can recognize a wide variety of intended actions, environments, or imagined stimuli from brain signals alone. This enables impressive new interfaces but still lacks precision compared to invasive BCIs.
Drawbacks are that EEG has low spatial resolution making the exact neural origin of signals unclear, and can be impacted by artifacts from facial/eye movements or nearby electronic devices. Emotions, attention, and fatigue can also impact recorded brain signals. Despite these challenges, non-invasive BCIs have the greatest potential for widespread usage due to accessibility and lack of medical risk factors.
Invasive Cortical and Depth Electrode Brain-Computer Interfaces
Invasive BCIs involve surgically implanting electrode arrays or probes directly into target regions of the brain or peripheral nerves. This provides a much higher resolution window into local neural circuits with improved signal-to-noise ratio compared to EEG. Surgery presents risks requiring careful clinical design and oversight.
Cortical electrode grids and microelectrode arrays are implanted directly on the brain surface or within cortical regions to record from neurons and neuron populations. Signals are processed to extract multi-dimensional tuning profiles of firing rates indicative of planned or imagined movements, sensory processing, or other internal processes.
Depth electrodes are long probes inserted into deeper structures like the hippocampus for epilepsy monitoring or basal ganglia for Parkinson’s disease therapies. They enable interfacing structures deeper than scalp EEG or cortical grid implants can access.
Invasive BCIs allow decoding of finer movement details like finger or limb positioning not feasible with non-invasive methods. By directly tapping neural coding, invasive interfaces achieve state-of-the-art accuracy in prosthetic device or cursor control. They have provided new assistive capabilities to paralyzed patients for communication, mobility, and control of devices through neural signals alone.
Besides surgical risks, chronic implants face challenges from the body’s response over time. Tissue scarring can degrade interface quality or require reoperation. Device failures may require repeated surgeries. Further research aims to minimize these risks through biocompatible materials, wireless power/data transmission, and tailored anti-inflammatory coatings on electrode shanks.
Brain Decoding Algorithms and Machine Learning
Brain decoding algorithms play a central role in Brain-Computer Interfaces by mapping recorded patterns of neural activity to intended actions, perceptions, or cognitive states. Early work focused on linear mappings of individual or small sets of neurons.
Modern algorithms leverage machine learning techniques to tackle increasingly complex, high-dimensional decoding problems posed by human brain activity. Deep neural networks are well suited to process the spatial and temporal patterns contained within invasive electrode array recordings or non-invasive EEG/fMRI signals.
Convolutional neural networks excel at exploiting the topological structure of brain regions and spatial tuning properties of localized neurons. Recurrent networks like LSTMs handle the dynamic, temporal development of neural responses and activity sequences over time.
Large datasets of paired brain recordings and behavioral/stimulus events enable training powerful supervised learning models. Reinforcement learning techniques optimize decoding when explicit targets are uncertain. Bayesian and hierarchical approaches capture uncertainty in neural population codes.
Ensemble and multitask learning methods outperform single models by combining evidence across diverse features, tasks, and models. Transfer learning exploits related tasks to boost performance in new contexts with fewer labeled examples. Real-time BCIs require low-latency incremental learning frameworks as well.
Machine intelligence augments human intelligence within brain interfaces. As brain datasets and computational power grow, automated online mapping of the brain at large scales will vastly enhance our understanding and ability to interface with the mind. This revolution hinges on advances in applied neuroscience, engineering, ethics and more.
Andrew Stroud