Neuralink’s activity was keeping all of us in suspense for the last two years.
“— What they invented? — Can we transfer our minds or uhmm… connect it to the WiFi hotspot?” – let’s put aside our futuristic fantasies for a moment, we will consider them later.
Recent Elon Musk’s announcement of Neuralink’s progress in brain-machine interfaces (BMIs) research was surprising with regards to number of improvements. However, there is a long long way to go until we will finally put some micro stuff into our skulls and control the environment or (*futuristic fantasy here*) read someone’s else mind.
In today’s post we will place this piece of research under scrutiny on the basis of Neuralink’s white paper  published on the day of Neuralink Launch Event. The goal is to present some highlights of Neuralink system, step by step, as it was presented in the paper cited. We will summarize by discussing the potential outcomes of proposed solutions.
Even if you’re familiar with advances in high-tech ideas, it would be worth asking: what is brain-machine interface in general? Well, the definition may seem broad, but BMIs are devices / sets of devices, which are between the brain and the machine (it can be a computer, a prosthetic limb, a module to switch something on/off – pretty much everything controllable). The goal of the BMI is to mediate between brain and machine by transferring the signal from one to another (may be in both ways).
What is the signal of BMI? The issue splits here on many levels. BMIs may be roughly divided into invasive and non-invasive. You may recognize such techniques as electroencephalography (EEG) or functional magnetic resonance imaging (fMRI) – these two are pretty good examples of measurement which can be used for non-invasive BMIs (non-invasive == no need for any surgical procedures). While using EEG, we record specific changes of electrical potential from the scalp; with fMRI we can investigate blood-oxygen level dependent (BOLD) responses – both signals can be recorded, processed and further used for steering purposes.
When it comes to invasive BMIs, it is needed to perform a surgery in order to put some stuff under the skull – it may be on the surface of the brain or within the cerebral cortex. Yes, you may feel the goosebump while imagining it. In the cerebral cortex there are many many cell bodies of neurons, and neurons are excitable cells, so they have electrical activity changing in time. And this is what we can basically record and use. And this is what Elon & Neuralink used for the purpose of their study.
Step-by-step – what’s inside Neuralink BMI?
The system consists of a surgical robot, implants and outer electronics to provide power and gather the data. Its modular structure is very promising, as it gives a chance to extend the system by placing the implants on the broad range of brain surface.
Let’s skim through every part of Neuralink’s BMI.
The robot – a.k.a the sewing machine
The surgical robot amazed me! It doesn’t look like the most complicated medical robot in the world, but possesses many features and is.. elegant. Only one part has direct contact with brain tissue – it is the needle with a thread, led by the inserter. The needle has its own linear motor, just to rapidly push and pull it from the cortex surface. What about other features? Wooh! Brain position sensors, six light sources of various wavelengths and cameras – together they play a huge role in localization of brain structures on the basis of known coordinates and depth of field tracking, but are also necessary for the neurosurgeon to monitor the procedure. The robot can work in auto mode (pretty impressive, 6 threads in one minute!), but hey, right now it is still okay to have an expert on your side. Especially when we talk about putting something in the brain tissue.
The stuff – from electrodes to USB-C
Let’s start from the electrodes. Authors wrote about their trials with two different materials for the electrodes themselves – PEDOT (a durable conductive polymer) and irydium oxide (they wrote about its better biocompatibility than PEDOT’s). Electrodes are placed on a polymer thread which is inserted into the brain tissue. The number of electrodes is impressive! On one thread there are 32 of them. Neuralink presented two systems – described as A and B – which consist of 1536 and 3072 electrodes respectively (Figure 2). Wow.
They invented their own, custom-made ASIC circuit (see Figure 2). It has 256 amplifiers, which can serve for the purpose of 8 threads in one time (8 threads — 256 electrodes, then 1 electrode — 1 amplifier). The rest of the circuit are ADC converter with sampling rate of ~ 19 kHz and circuits to pack and send the data through the USB-C. USB-C also provides power for the system, of about 6 mW per one ASIC.
What was recorded so far?
One of the goals of this post was to briefly explain what was recorded during Neuralink’s experiments, because some concepts from neurophysiology were not defined in the published white paper. Authors wrote that the signal of their interest comes in two: spikes and local field potentials (LFPs). What are those?
- Spikes (more formal: action potentials) are rapid changes in polarity of e.g. neuron’s membrane. As you know, each cell has a boundary – in the case of animals it is a membrane consisting of various proteins, phospholipids etc. which selectively transmits products inside or outside the cell. The outer and inner environments of cells differ with regard to ions distribution – if it is stable we say that it is the resting potential. When it comes to neurons, if channels on a membrane open (because of being excited) there are rapid changes happening in the cell resulting in depolarisation or hyperpolarisation. If it accumulates and reaches a threshold, a spike is triggered. 
- Local field potential (LFP) is recorded from the group of nerve cells in the neighbourhood of the electrode. It is a sum of electrical activity of bunch of neurons, although the source of the electrical activity recorded lies not in individual action potential, but in synaptic and dendritic currents. 
Well, as you can see, it can be pretty difficult to see consciousness or secrets of our minds in such signals, huh? Authors didn’t provide the approximate coverage (in the number of neurons) of one array of electrodes – that would be interesting to see how it may scale up. I realize that the final coverage is task-dependent – if one needs to stimulate a human motor cortex it could be way different from the stimulation of e.g. Wernicke’s area, as these two brain regions differ in size.
Neuralink used this approach to study their BMIs on rats, while they freely explored the space. Authors were recording the signals with the usage of systems A and B with online detection of some structures (such as spikes). The sewing robot performed 19 surgeries on rats with a success rate of 87%, although it was not described how the performance was measured (succeeded depth of insertion? errors of the robot? thread breaking?).
During the Neuralink Launch Event there was a presentation of future ideas for the application of BMI in humans. The general idea behind it is that the system will consist of implants (similar to those presented in paper cited, ), which will be connected to an external wearable device looking like a hearing aid. This wearable will connect via Bluetooth with Neuralink app on your iPhone. This is cool. But there are also many doubts arising? What about data security? What about data storage? The amount of data from the brain will be probably huge and not useful as a whole, but some specific tendencies or information about one’s brain / mental health is something. Something we’d like to protect.
- A lot and a little. There is a lot of improvement done. Thread with electrodes seems to be a fast and robust solution while considering some future clinical applications. There is also an innovative system of its application – as I wrote above, speed of the robot and its additional features are together impressive. But still, there is a lot to do if we want to obtain a fully BMI – in particular with the interpretation of extracellular signal, as we need to assign it to specific tasks done by machine.
- Modulation of the neurons. I like it. I didn’t mention it earlier, but Neuralink’s team claim that the electrodes will be able both to record and stimulate brain tissue. It opens a wide range of possibilites, especially for the purpose of disorders related to changes in concentration of specific substances. A closed-loop cybernetic system.
- A lot of channels and a modular system. This is a huge advantage, as one can design the displacement of electrodes for particular clinical or non-clinical purposes (in the future:-)).
First I read the paper. Afterwards, I decided to watch Neuralink Launch video. In the paper the future promises given by Neuralink’s improvements in BMI research are restricted to one paragraph, with only a brief comment on future application in humans for the purpose of neuroprosthetics and restoration of motor function. On the other hand, there are some promises in the Launch video, which today are still beyond the reach. For example, schizophrenia and depression were mentioned, but currently there are no well-established models of these disorders, which could be used to design treatment via stimulation with feedback loop (if there are, please DM me, but so far I didn’t see such).
I somehow understand the difference in the narrative – these futuristic (but reachable!) ideas need to be presented in such way, because young minds have to be infected with the idea that this is very, very possible way of treating some of our current problems. I agree with the statement “The Future is Now” and I love it, even if there is one year or two to go.
 Musk, E., Neuralink (2019) An Integrated Brain-Machine Interface Platform with Thousands of Channels. (white paper)