Neuralink, founded by Silicon Valley iron man Elon Musk, has been a hot topic since it came into public view in 2016. Recently, Elon Musk, the founder of neuralink, has made a new move.
on July 20, Elon was asked by computer scientist Austin Howard on Twitter: can neuralink technology realize listening to music directly from a chip? The Silicon Valley iron man replied: Yes!
this answer has made many people who don’t know much about BCI instantly fill a lot of pictures. In the matrix, which was released in 1999, through the invasive brain computer interface and brain nerve connection, human beings can feel the signals of vision, hearing, smell and taste, which can imprison the human mind.
in “X-Men”, Professor X can amplify brain waves through “brain wave intensifier” to connect with anyone. In the movie Avatar, the protagonist, through the combination of EEG and EMG, can control the body of the neville people and move on the planet Pandora.
neuralink tells you that human experiments may be carried out by the end of 2020, which will make people feel that BCI is very close to our lives.
so, at this stage, is brain computer interface reliable? What stage has it progressed to and what challenges will it face? Through this article, we will talk about BCI, a lot of things you don’t understand.
the so-called brain computer interface (BCI) refers to the direct connection between human or animal brain (or brain cell culture) and external devices. Among them, “brain” refers to the brain or nervous system of organic life form, “machine” refers to any processing or computing equipment, its form can be from simple circuit to silicon chip to external equipment and wheelchair, “interface” = “intermediary for information exchange”. Therefore, the definition of “brain computer interface” is “brain” + “machine” + “interface”.
the basic implementation steps of BCI can be divided into four steps: single acquisition – Feature Extraction – frequency translation – feedback. Next, we will explain the details of each process one by one according to this process, so as to understand the difficulties of BCI at present.
if we make an anatomic plane of the human brain, the order from inside to outside is: scalp (scalp), skull (cranium), dura mater (dura mater), arachnoid (arachnoid), pia mater (pia mater), and cerebral cortex (cerebral cortex).
invasive BCI refers to a device that implants sensors into brain tissue to obtain signals through craniotomy. Its disadvantage is easy to induce immune response and callus, and then lead to the decline or even disappearance of signal quality. This is how Elon Musk’s neuralink works.
semi invasive BCI is a device placed on the surface of the cerebral cortex to receive signals. The interface is generally implanted into the cranial cavity, but it is located outside the gray matter. Its spatial resolution is not as good as that of the invasive BCI, but better than that of the non-invasive BCI. The advantage is that the probability of inducing immune response and callus is small, and the information analysis is mainly based on the cortical electroencephalogram (ECoG).
non invasive BCI, which does not enter the brain, i.e. a device that detects signals outside the skull. This form is as easy to wear as a hat. However, due to the attenuation effect of skull on the signal and the dispersion and blur effect of electromagnetic wave on neurons, the resolution of recorded signal is not high, so it is difficult to determine the brain area or the discharge of related single neuron. The typical system is EEG (electroencephalogram). The advantages of EEG are its good time resolution, ease of use, portability and relatively low price. However, one of the core problems of EEG is its poor sensitivity to noise. In addition, using EEG as brain computer interface requires users to carry out a lot of training before using, in order to better operate the non-invasive brain computer interface.
the higher the degree of invasion, the higher the signal quality and strength obtained, and the higher the risk. According to the signal quality: intrusive semi intrusive non intrusive. By risk: intrusive, semi intrusive, non intrusive.
after collecting enough information, the signal must be decoded and recoded to deal with the interference. There are many interferences in the process of EEG signal acquisition, such as power frequency interference, eye movement artifact and other electromagnetic interference in the environment.
the analysis model is the key to information decoding. According to the different collection methods, there are usually models such as EEG and ECoG to assist in the analysis.
the methods of signal processing, analysis and feature extraction include denoising and filtering, P300 signal analysis, wavelet analysis + singular value decomposition, etc.
encode the analyzed information. How to code depends on what you want to do. For example, to control the robot arm to pick up a coffee cup for oneself to drink coffee, it needs to be encoded into the motion signal of the manipulator. It is very complex to accurately control the movement trajectory and force control of the object in the complex three-dimensional environment.
but the coding forms can also be various, which is why BCI can be combined with almost any engineering discipline. The most complex scenario involves exporting to other organisms, such as mice, to control their behavior.
it is also very complicated to act on the brain after obtaining environmental feedback. Human beings perceive the environment through the ability of perception and transmit it to the brain for feedback. Perception includes vision, touch and hearing.
the implementation of BCI is very complicated, including the mixed analysis of multimodal perception, which may be incompatible with the process of feedback to the brain.
we can first take a look at the technology of neuralink, which is at the leading level, to see the current development.
the paper written by CARMENA in 2003 is the core thesis in the field of invasive brain computer interface, which lays the foundation for this field. Let’s compare the “classic” method proposed in 2003 with Elon on the neuralink website The invasive brain computer interface technology of neuralink mentioned by musk is the first author. It is found that the improvement core of neuralink is mainly in three aspects:
using thread instead of electrode: the needle size is smaller, the number is more, and the damage to the brain is smaller, so the signal-to-noise ratio and signal source can be extracted. It is more efficient to use a robot for electrode installation. A special chip (ASIC) is used to preprocess the signal and extract the feature of the signal
. The so-called thread is a smile displacement nerve probe made of a variety of biocompatible thin film materials. Compared with the rigid metal or the electrode array of this semiconductor process, this material has higher biocompatibility and is not easy to cause immune reaction caused by the mismatch between Young’s modulus and bending stiffness.
at present, neuralink has constructed small and flexible electrode “needle” arrays, with up to 96 stitches per array and up to 3072 electrodes distributed.
because the film material used to make the needle is very thin and not hard enough to be implanted into the brain, neuralink has developed robot technology to perform implantation surgery.
the robot can insert six needles (192 electrodes) per minute. Each needle can be individually inserted into the brain with micron accuracy to avoid the surface vascular system and target specific brain regions. The electrode array is packaged in a small implantable device that contains custom chips for low power on-board amplification and Digitization: the area of the package for 3072 channels is less than (23 × 18.5 × 2) mm3.
high density recording channel requires that signal amplification and digital to analog conversion must be integrated into the array components. Moreover, the integrated module must be able to amplify weak neural signals (10 μ VRMs) and suppress noise. At the minimum power consumption and size, the amplified signals are sampled and digitized, and processed in real time.
neuralink’s ASIC can meet the above requirements. The integrated circuit consists of three parts: 256 independent programmable amplifiers (analog pixel as neuralink calls it), on-chip analog-to-digital converter (ADC), and peripheral control circuit for serializing digital output.
although neuralink has made some technical optimization and breakthrough, both the scheme proposed in 2003 and the current neuralink rely on the feedback closed loop of “brain machine” and “machine brain”. In other words, the essence is to solve the following two problems:
the second is “from machine to brain”, which inputs information into the brain or changes the natural flow of the brain in other ways – this is stimulating neurons.
the existing brain computer interface technology is only a preliminary solution to the problem of “brain computer” output and control, but the control efficiency and accuracy are very low. This is because of the limitation of the basic principle, it is necessary to reconstruct the existing brain computer interface technology fundamentally, otherwise, the potential of this technology will be difficult to tap quickly.
compared with the problems in the “brain machine” direction, the “machine brain” aspect has to face more difficulties, with almost no clue and only a few lights.
what does “brain machine” mean? That is, the perception is encoded in reverse into signals that can be read by the brain. For example, it’s also a good idea to reconstruct the vision of the blind by using the machine to reproduce the touch or a piece of your imagination when you touch a kitten.
the research on “machine brain” is much slower than that of “brain machine”. The reason is that the specific way of neural coding in neuroscience is still unknown. However, the demand for neural coding knowledge from machine brain is far greater than that from brain machine. Neuroscience in the study of single neuron is also gradually clear, but the various magic of the brain simply can not explain.
according to the chart above, at the rate of doubling the number of neurons that can be recorded at the same time in an average of 7.4 years, it will take 2100 to record one million neurons at the same time, and 2225 to record all neurons (5-10 billion) in the human brain. Therefore, how to solve the bandwidth problem of BCI has become another key problem of academic research breakthrough.
brain computer interface is an interdisciplinary research field. The core disciplines involve physics, machinery, neural engineering, electrical engineering, neuroscience, etc., which are more practical than engineering. Multidisciplinary is just the theoretical basis of engineering implementation.
at present, the development of BCI needs the joint development of multiple disciplines. The backwardness of any discipline will affect the overall development process.
for example, the development of physics provides theoretical knowledge support, explains the problems in product testing from the perspective of principle, and solves the design and application problems of underlying sensor principle from the theoretical level. Neural engineering, especially experimental neurology and clinical neurology, provides theoretical support for the brain mechanism of brain computer interface and corresponding support products