Brain-Computer Interface and AI: Two Sides of the Same Coin
AI and neuroscience are like two sides of a bridge. They are moving towards each other and will eventually meet in the middle.
01
Neuroscience has inspired many key breakthroughs in AI.
Brain-Computer Interface (BCI) and AI may seem unconnected at first glance, but they have high consistency at the underlying logic level and share a compatible past and common future. Looking back at the past 20 years of AI, especially in the last 3 to 5 years, neuroscience has played an important role in several key technological breakthroughs of AI. For example, the design of neural networks comes from the study of the principle of brain neuron firing. Image recognition stems from the study of how visual neurons process visual signals. The Transformer architecture and the attention mechanism behind large language models come from the study of human attention. Jeff Hinton, who won the Nobel Prize this year, once said that the entire research of artificial intelligence is based on the foundation of neuroscience, but he also feels that the research on neuroscience is too slow now.
The brain and GPU have some similarities. Taking Nvidia’s H200 as an example, it has 10 to the power of 11 transistors. The brain has about 86 billion neurons, which is also roughly 10 to the power of 11. However, there are significant differences in the connections between transistors in a GPU and those in the brain. The connections between brain neurons can reach 10 to the power of 4, while the connections between transistors may be in single digits. In addition, the brain is a combination of storage and computation, and software and hardware cannot be separated. In computers, storage and computation are separated, and software and hardware can be separated. Some people ask whether AI can quickly simulate the entire brain. The author believes that there is probably still a difference of 10 to the power of 9. Part of this difference comes from the fact that the connection pathways between neurons are much more complex than those between transistors. Another part comes from neural plasticity, that is, the connections between neurons can change dynamically, while the connections between transistors cannot.
02
What is the brain-computer interface industry doing?
In neuroscience, the brain is the most important and fragile organ. At present, humans may understand less than 10% of the brain. We know some basic functions of different regions of the brain, such as which areas manage movement, language, hearing, etc., but the locations of more advanced functions such as knowledge, memory, emotion, consciousness, and self-awareness are still unclear. In the definition of neuroscience, all human perception, cognition, memory, emotion, and mood are determined by electrical signals. As long as electrical signals can be read, there is an opportunity to modify them.
The essence of a brain-computer interface is to insert electrodes into different regions of the brain. The electrodes, like wires, read the firing signals of neurons and can also stimulate neurons. Brain-computer interface is a typical interdisciplinary field involving electrode materials, chip packaging, medical devices, algorithms, etc., and it is extremely difficult. Its purpose is to make the brain benefit more and suffer less trauma when implanted with a brain-computer interface.
The brain-computer interface industry mainly has three technical routes. One is hard electrodes, such as a silicon chip the size of a fingernail with 100 needles that can be directly inserted into the cerebral cortex to record the firing signals of 100 to 200 neurons. One is a vascular stent that is placed in the brain through a vein and records neurons through the blood vessel. Another is a flexible electrode system, such as the one used by Musk’s Neuralink and Brain Tiger Technology. Sixty people worldwide have been implanted with the first technical route, ten with the second route, and three with the third route.
There are patients with high paraplegia who can control a robotic arm to cut cake and eat after being implanted with a brain-computer interface. Although the current effect can only be achieved in the laboratory. The vascular stent route has a lower ceiling. Neuralink’s flexible electrode system has made new progress this year, making the device extremely small and wireless. After patients are implanted, they can fully control a computer to play games, and the speed is even faster than that of many healthy people.
The development of brain-computer interfaces requires three core elements: high throughput, low trauma, and long-term presence in the body. Currently, Musk can record 1024 neurons, while Brain Tiger Technology can record 256 neurons, so there is still a gap. However, there is a Moore’s Law for brain-computer interfaces, that is, the number of neurons that can be read and written doubles every 18 months.
03
Why is the development of neural technology slower than that of AI
