Bright Pictures / Visual China
Bright Pictures / Visual China
Bright Pictures / Visual China
Bright Pictures / Visual China
Editor's note
With limited size and very low energy consumption, the human brain can complete cognitive tasks such as information association memory, rapid recognition, and self-learning Xi in complex environments. With the emergence of challenges such as the slowdown in chip replacement and the shortage of computing power, the computing methods of existing computers are no longer sustainable for the future. How to promote the further development of information processing technology and create a "machine brain" like the human brain? Brain-like computing is the direction to break the situation. In order to realize brain-like computing, it is necessary to find more neuromorphic devices that simulate the function of the human brain.
Learn from the human brain to make computing more efficient, more bionic, and less energy efficient
The invention of the electronic computer brought human civilization into a digital technological revolution. In just a few decades, the computing power of computers has increased by 1,016 times from hundreds of operations per second to 10 exascale operations per second, creating the highest artificial growth rate in the history of human civilization for thousands of years. The increase in computing power is due to the increase in the integration density of the underlying semiconductor devices (silicon-based transistors). According to Moore's Law, the number of transistors on a single chip doubles every 18 to 24 months.
For more than half a century, the integrated circuit industry has been developing under the guidance of Moore's Law. However, the reality is that as Moore's Law continues to approach its limit, the transistors that a single silicon-based chip can carry are becoming more and more saturated. The size of the silicon atom is about 0.12 nanometers, and when the chip process reaches 1 nanometer, there is only room for a few silicon atoms to be manipulated.
In fact, after the chip process is developed to the 10nm level, problems such as the slowdown of replacement speed and rising costs can be clearly felt. On the other hand, as the wave of artificial intelligence models sweeps the world, the demand for computer computing power has surged to the extent that it will double every 2 to 3 months, far exceeding the growth rate under Moore's Law.
In the face of challenges such as the slowdown of chip replacement and the shortage of computing power, the computing method of digital computers has become unsustainable, and how to promote the further development of information processing technology has become a common problem faced by academia and all walks of life. In the face of this huge dilemma, brain-like computing that can provide more efficient, more bionic, and lower energy computing power has become the key to breaking the game.
Brain-inspired computing is an important research direction in the International Semiconductor Technology Blueprint (ITRS/IRDS), which aims to draw on the basic principles of the human brain to realize artificial general intelligence (also known as brain-like general intelligence). In the "China Brain Project" launched in 2021, brain-inspired computing is an important component.
Different from traditional computers, the human brain is able to complete cognitive tasks such as information association memory, rapid recognition, and self-learning Xi in complex environments with limited size and extremely low energy consumption. This is closely related to the basic composition and structure of the human brain's neural network: there are 86 billion neurons in the human brain, equivalent to the number of celestial bodies in the Milky Way, and 150 trillion neural synapses are interconnected to form a spatially complex neural network, while the neural dendrites and other tissues of the human brain further complicate the neural computing function.
These neural tissues of the human brain contain a variety of ion channels, have very rich dynamic behaviors, and feature time scales spanning several orders of magnitude, which are the physical basis of human brain intelligence. Correspondingly, the basic building blocks of a computer are electronic transistors, which operate in quasi-static 0 and 1 encoding, which is far from the rich dynamics of the human brain.
Therefore, one of the keys to the realization of brain-inspired computing is the discovery of neuromorphic devices. They can simulate the functions of neurons, synapses, and nerve dendrites in the human brain, and have a physical mechanism that is closer to the behavior of neural tissues, so that many kinds of neurological functions can not be achieved by traditional electronic transistors.
Brain-inspired computing based on neuromorphic devices is developing rapidly
Neuroscience research has found that the modulability of the strength of synaptic connections between neurons is one of the foundations of brain Xi and memory function. Changes in the strength of synaptic connections caused by past experiences can have an impact on brain function.
Changes in the strength of synaptic connections, also known as synaptic plasticity, can enhance or inhibit neuronal activity and can last from milliseconds to hours, days, or even longer.
If we can learn from the principle of synaptic plasticity, imitate and realize it by some means, build artificial synapses similar to synapses, and then further build a system, we can better understand and simulate the way the brain works, further promote the cross-development of informatics and neuroscience, and realize brain-like computing.
As early as 1971, scientist Cai Shaotang had heuristically reasoned and predicted a new type of device, the memristor. According to the prophecy, the resistance value of the memristor depends on the applied voltage/current excitation history and therefore has neuromemory-like properties.
Thirty-seven years after this prediction, Hewlett-Packard Labs announced that the memristive phenomenon has been observed in new micro/nano semiconductor devices. Since then, memristive devices and neuromorphic devices have become almost two interchangeable concepts, and brain-like computing based on neuromorphic devices has also entered a stage of rapid development.
As a potential circuit element, memristors are superior to traditional transistor devices in terms of scalability, memory density, and power consumption in addition to biological similarity.
In recent years, important progress has been made in both materials technology and function, and neuromorphic devices. In terms of materials technology, researchers have applied a wide range of materials—inorganic, organic, quantum, ferroelectric, ferromagnetic, three-dimensional, and two-dimensional materials—which exhibit their own unique neuromorphic properties, providing diversity and flexibility for the development of memristors. Significant progress has also been made in the research of neuromorphic integrated circuits that combine traditional transistors and memristors, which has accelerated the application and popularization of memristors. In terms of function, memristors can not only simulate the plastic functions of neural synapses, but also some functions of neurons, which creates the possibility of realizing the neuromorphic circuitry of full memristors.
The "non-ideal" physical mechanism of transistors is used to simulate the memory function of the human brain
However, the development of neuromorphic devices at this stage presents new challenges. One of the key challenges is that biomimetic dynamics are under-functional to meet the requirements of brain-inspired computing for rich neuromorphic dynamics.
As mentioned earlier, the rich dynamical behavior of the human brain is closely related to the diverse structure and mechanism of ion channels in neural tissue. However, at present, mainstream neuromorphic devices are usually customized to simulate a specific neural behavior, using a specific single physical mechanism to achieve it.
If rich biomimetic dynamics are to be realized, full-featured kinetic neuromorphic devices need to be developed. However, in general, the more versatile the features, the larger the size of the hardware, which contradicts the miniaturization of current chips. To solve this problem, it is necessary to explore new device principles and new semiconductor materials.
As mentioned earlier, one of the major features of synaptic plasticity is that the dynamic time scale spans several orders of magnitude, which is a basis for human cognitive and memory function. In fact, we can all feel this dynamic – sometimes one thing is unforgettable, and sometimes the next is forgotten. This is the dynamic behavior of long-term memory and short-term memory on different time scales, and their coexistence helps us retain important information while filtering out unimportant information, reducing the burden on the brain. However, the existing single-device artificial synapses can only be selectively simulated for long-term plasticity or short-term plasticity, and cannot be integrated for simulation.
Based on these principles of neurosynapses, a comparison of artificial synaptic devices and biological synapses reveals a huge difference between them – the former uses the same physical mechanism to simulate both functions, while the latter utilizes different calcium channel mechanisms from the postsynaptic and anterior membranes, respectively.
Inspired by this, the research team at Tsinghua University's Brain-like Research Center turned their attention back to transistors. As the basic components of computer chips, transistor devices actually contain two physical mechanisms - the "field effect" mechanism and the "memristor" mechanism. The "field effect" mechanism allows the transistor to switch between the 0 and 1 states, but without continuous power, the state will quickly disappear, which is unsatisfactory from the perspective of energy efficiency. The memristor mechanism confuses the 0 and 1 states, which has been seen as a detrimental effect in the past, so it is important to prevent the emergence of the memristive mechanism when manufacturing computer chips. But the memristor mechanism has another property – it can persist even after a power outage.
Aren't these two physical mechanisms, which are imperfect or even unfavorable for traditional computer chips, exactly what are needed to simulate the long-term and short-term plasticity of synapses in brain-like computing?
At this point, the answer is coming. Through the "reverse" use of these two mechanisms, the dynamic neuromorphic transistor technology proposed by the team of the Brain-inspired Computing Research Center of Tsinghua University enables the dynamic functions of long and short-term memory to be simulated in a single device, which solves a key technical problem in the field of brain-inspired science.
Find more semiconductor devices that can simulate neural computing in the human brain
Neurons are another basic neural computing unit whose suprathreshold firing and subthreshold oscillations are involved in almost all cognitive functions, essentially the rise and fall of cell membrane potentials. From a biological point of view, changes in neuronal membrane potential are involved in two ion channels – sodium (Na+) channels and potassium (K+) channels.
They work like this – when the sodium channel is opened, the extramembrane sodium (Na+) flows inward into the cell, resulting in an increase in the membrane potential, which is called the depolarization process, and after the membrane potential reaches a certain level, the repeated polarization process begins, the sodium (Na+) channel closes, and the potassium (K+) channel opens, allowing the potassium (K+) to flow out of the cell, thus reducing the membrane potential.
The dynamics of neurons are more complex, and to model them often requires the combination of multiple electronic components into circuits. In order to make artificial neurons both dynamic and hardware simple, it is necessary to find new materials to achieve this.
Eventually, tellurium, a new semiconductor material, came to the fore. It has a combination of physical properties such as low melting point, low thermal conductivity, and electrochemical activity, which are difficult to find in other materials.
These properties allow the tellurium conductive channel structure to grow under the electric field of the electric current, reducing the resistance of the device, which can correspond to the inflow depolarization process of sodium ions (Na+), and the subsequent current Joule heating will fuse the conductive channel of tellurium and restore the resistance of the device, which corresponds to the outflow repolarization process of potassium ions (K+).
On this basis, the Tellurium semiconductor single device developed by the team of the Brain-like Computing Research Center of Tsinghua University has realized the full-function simulation of suprathreshold discharge and subthreshold oscillation of neurons.
Compared with neurons and synapses, nerve dendrites, as a typical characteristic structure of biological neural networks, have been compared to simple wires. However, more and more studies have shown that neural dendrites have important neural computing functions, not only performing passive calculations, but also actively "discharging", which may be one of the key sources of the human brain's general intelligent information processing ability and an important source of inspiration for enabling brain-like computing.
The active firing dynamics behavior of neural dendrites also stems from the abundance of ion channels. It is mediated by calcium ions (Ca2+), the duration of the discharge is longer, the effect may be more significant, and its activation function can also show a non-monotonic response to the intensity of the input stimulus. This allows a single dendrites to solve more challenging nonlinear classification problems.
In order to simulate neural dendrites, the team from the Brain-like Computing Research Center of Tsinghua University adopted a novel transistor structure, that is, a transistor based on the PN heterojunction semiconductor channel, instead of the traditional homogeneous or homogeneous semiconductor channel, and used the unusual "anti-bipolar" transfer characteristics of this special transistor structure to simulate the non-monotonic activation and dendritic discharge mediated by calcium ions (Ca2+), further enriching the neuromorphic function of the transistor.
Today, although some progress has been made in simulating the neural calculations of the human brain, the brain, as a collection of human intelligence, is the most complex product of the known universe, and the study of the brain is also called the "ultimate frontier" of natural science. As a machine intelligent computing that mimics the mechanism of neurophysiology and physiological psychology, uses computational modeling as a means and realizes the collaboration of software and hardware, brain-inspired computing is still a challenging way to go before realizing the dream of human beings to build a "machine brain" like the human brain.
Source: Guangming Daily
Statement: The content is transferred from the Internet, the copyright belongs to the original author, if there is any infringement, please contact to delete.