3D, Three-dimensional

BCI, Brain-Computer Interface

BMI, Brain-Machine Interface

CNS, Central Nervous System

CPU, Central Processing Unit

DPU, Decoding Processing Unit

ECoG, Electrocorticography

EEG, Electroencephalogram

fMRI, Functional Magnetic Resonance Imaging

PET, Positron Emission Tomography

PNS, Peripheral Nervous System

TPU, Tensor Processing Unit

In the last year, there has been an explosion of interest by entrepreneurs looking to become actively involved in developing devices to interact with the central nervous system. These have included the likes of Elon Musk (Neuralink Inc. California USA), Mark Zuckerberg (Facebook Inc. California, USA), Bryan Johnson (Kernel. California, USA) as well as dedicated startups such as Paradromics (San Jose, California, USA) or Cortera (Berkeley, California, USA), and even DARPA (Defense Advanced Research Projects Agency. Virginia, USA), spurred on in part by the BRAIN initiative ¹ . Each of these individuals and their respective companies share a particular focus in developing a robust brain-computer interface (BCI). We define BCI, for the purposes of this discussion, as a technological system designed to provide a stable mapping and modulation of activity within neural networks of the central nervous system. Therefore, at the very minimum, a working BCI will require both a physical interface to the brain (brain-machine interface; BMI) and computer systems that can process high bandwidth signals in real-time.

It is important to distinguish that there are very different engineering and neurological challenges between building BCIs for the peripheral nervous system (PNS) and central nervous system (CNS). In particular, space limitations for processing units, isolation of targeted structures, replacement of probes following failure, and delivery of nanomaterials in vivo ^{2, 3} ; for the purpose of this commentary we will focus on the CNS as this is an area of particular interest by the entrepreneurs highlighted above.

Understanding the information transfer and processing of the nervous system is one of the most urgent challenges faced by the biomedical community, with a plethora of academic and clinical applications, including better understanding of aging, neurodegenerative diseases and interfaces for prosthetics and implants. For example, recent advances in chronic neural recording devices have facilitated the willful control of robotic prosthetic limbs for the treatment of paralysis ⁴ and improved seizure prevention with chronic telemetry in refractory epilepsy ^{5, 6} . There are many different kinds of potential BCIs that will each serve independent functions, however all systems must tackle three fundamental problems: how to accurately record information from relevant neural systems, how to decode such information, and how to stimulate and manipulate neuronal dynamics in an appropriate and meaningful way.

The origins of neural engineering stretch back to early attempts to record activity chronically in the 1950s when electrodes were implanted into the cortex of rhesus monkeys to measure electrical activity in the central nervous system ^{7, 8} . Great innovations have been made in neural recording techniques, which have allowed the number of simultaneously recorded neurons to double approximately every 7 years ⁹ , mimicking Moore’s law albeit at a much reduced rate ¹⁰ . Early clinical applications of BMIs centered on the restoration of perceptions to patients with sensory deficits. One of the pioneering studies was the work on potential cochlear implants in the 1970s that eventually reached life-changing reality in the 1980s for patients ^{11– 13} .

In parallel to the development of the cochlear implant, researchers worked with the CNS by applying electrical current to the visual cortex of blind patients through grids of surface electrodes implanted over the visual cortex, thus developing visual prostheses ^{14, 15} . These systems allowed blind subjects to learn to recognize simple visual objects ¹⁶ . Neural engineering continued to improve with multi-channel neuronal recordings allowing owl monkeys ¹⁷ and later humans ⁴ to control two- and three-dimensional movements of a robot arm with multiple degree of freedom. Neuro-prosthetic research has undoubtedly benefited from these advances, but additional design parameters need to be included for effective long-term operation and clinical translation of neural interfaces.

While research in neural engineering has been steadily improving the bandwidth of BCI interfaces, the pace of this exponential increase falls far short of that seen in the silicon chip industry ⁹ . At current pace, the goal set by DARPA of recording from 10 ⁶ neurons simultaneously would not be expected to be reached for around 80–100 years. Increasing interest and funding from members of Silicon Valley may prove to be a useful catalyst for the field and promote investigation of new applications of BCIs. For example, Facebook Inc. is investigating methods of non-verbal communication that will not require the virtual keyboards that are currently being used by patients with BrainGate ¹⁸ .

Despite advances in recent years, implementation of general purpose BCIs faces a number of constraints arising from engineering, computational, ethical, and neuroscientific factors. The future success of BCI is often imagined as a function of the capability to produce multi-electrode arrays with a greater and greater density of recording sites. Here, we outline several other challenges that must be overcome in parallel if BCI is to become of more than limited interest.

Perhaps the most immediate barrier to wider usage of BCI systems is the difficulty in implanting them. Non-invasive modalities, such as electroencephalogram (EEG) but also positron emission tomography (PET) and functional magnetic resonance imaging (fMRI) lack the spatial resolution to record detailed activity at the level of the neuronal circuit, and so can only be used for very simple low bandwidth (typically binary choice) interfaces. There is no technology currently available that can record an action potential without the need for major surgery, although research into less invasive endovascular electrodes ¹⁹ and surface electrocorticogram (ECoG) devices is ongoing ²⁰ . Furthermore, the quality of recordings obtained via implantable electrodes degrades over time due to a combination of gliosis ^{21, 22} , neuronal depletion ^{23, 24} , and degradation of the system itself ^{22, 25, 26} . This tends to limit recording times to a period of months or a few years at most, although the use of compliant materials ²⁷ or soft ultra-thin wires ^{28, 29} designed to reduce mechanical shear has shown promise in reducing these effects.

By definition, detecting neuronal signals constitutes only one half of the BCI. These signals must then be able to be communicated to a computer via either a wired or wireless connection. This poses further challenges, necessitating a tunneled wire through the cranium. Wireless systems avoid this challenge, but create a host of new problems in turn including available bandwidth, safety, and the need for an implantable battery – which may last only a few months powering a large BCI system ^{30, 31} . To give a sense of the challenge here, we calculate that a 100,000-electrode system would require a communication protocol at least as fast as a Thunderbolt ^TM 3 connection (Apple, Inc. & Intel, Inc.), currently the fastest available consumer-level wired standard. The required bandwidth could be reduced drastically by on-chip processing, reducing the dimensionality of the data, but this in turn requires vastly more complex devices, limiting the number of electrodes per device, and greatly increasing its volume – a critical flaw in any proposed intracranial device. Furthermore, onboard processing of any kind poses serious and mostly unexplored challenges in terms of the energy dissipation required to maintain the device at body temperature so as not to cause thermal damage to the brain.

Current multi-electrode array systems offer up to around one thousand recording channels ³² , in turn providing monitoring for hundreds of neurons from a single area ³³ , sufficient for the control of several univariate parameters. More general purpose BCI will require the sampling of tens if not hundreds of thousands of units, potentially from multiple cortical regions. This poses engineering and surgical challenges far beyond what is currently achievable.

Computational and data analysis challenges arise from the highly parallel nature of multiunit recordings. In general, there are four steps utilized to decode neural activity. Firstly, the signal must be filtered to remove extraneous noise. Secondly, spikes must be detected. Thirdly, these spikes must be ‘sorted’, typically by waveform, in order to be assigned to ‘units’ – putative single neurons. Lastly, the inferred population spike train must be decoded in order to provide a control signal. Whilst the first and second of these steps are essentially solved, for sufficiently high signal-to-noise systems, spike sorting is still an area of active research ^{34, 35} , with no clear optimal solution, and often relies on semi-automated systems that require a great deal of human input to fine tune. Spike sorting may not be strictly necessary for the training of accurate decoders, as the raw spatiotemporal pattern of activity may suffice, but this may in turn reduce the dimensionality of the data.

Real-time processing of highly parallel recording systems remains a key challenge in the field. Promising technologies include a move away from general-purpose central processing units (CPUs) to application specific integrated circuits designed to perform a limited number of operations, such as Google’s tensor processing unit (TPU) or the graphical processing chips found in most computers. It is not unreasonable to suspect that the solution to decoding neural activity may lie in dedicated ‘decoding processing units’ (DPUs).

The physical scalability of BCI systems also poses a profound challenge. The brain is a three-dimensional (3D) structure. Unlike silicon wafers, manufacturing devices with a complex 3D structure and including integrated electronics poses a particular problem. Furthermore, current designs of multi-electrode arrays are typically not well suited to rapid scalability, requiring extensive redesign for each generation of device.

Even if this problem can be overcome, it may seem intuitive that more units result in greater bandwidth, however, the distributed nature of cortical processing has actually shown to result in a decreasing marginal value of each additional unit in terms of information retrieval ³⁶ . Therefore, the common mantra that more units results in more information does not follow, at least not proportionally. We simply do not understand well enough the nature of distributed information representation and processing in the neocortex to be able to make more than a rudimentary estimate of what a particular sequence of activity might ‘mean’.

The literature has shown large decades of neuroscience research efforts in developing tools to probe the signaling complexity of the nervous system, with several clinical applications being developed. Although orders of magnitude smaller and computationally faster than neurons, our electronics cannot mimic the complexity of neural systems. Current understanding of the function of neural circuits could be compared to trying to understand the internet by means of a few dozen well-placed potentiometers in the data centers of service providers. This is not to disparage the efforts of neuroscientists, far from it, but rather to underscore that decoding neural circuits ranks among the deepest and most complex contemporary endeavors, and it will not be solved overnight by Silicon Valley enthusiasm and zeal alone. However, we consider that many of the engineering challenges outlines above are amenable to focused research and development, particularly those surrounding miniaturization and parallelization of recording systems. We support the interest of entrepreneurs in placing their focus on the neuroscience community, and we look forward to the future advancements that will undoubtedly be realized in the coming years.

No data are associated with this article