Release date: 2014-09-01
In 1982, Kwabena Boahen had his first computer, when he was only a teenager who lived in Accra, Ghana. "That is really a very cool device," he recalls. But when Bo Heng figured out how the internal workings of the computer, he didn't feel particularly shocked. "I figured out how the central processor constantly transfers data back and forth. I thought, 'Scorpio! The computer needs to work madly all the time'." He instinctively felt that the computer needed more "African" features in the design. The design: the distribution is more extensive, more fluid, less strict.
Today, as a bioengineer at Stanford University in the United States, Bo Heng and other researchers have formed a small team that is trying to create the ideal computer operating model by mimicking the brain.
The energy savings of the brain are significant, and its computing power is enough to challenge the world's largest supercomputer, although it relies on components that are not perfect: neurons are very slow, variable, and confusing. In a smaller area than the shoebox, the brain can complete the understanding of language, abstract reasoning, control motion and even more tasks, but the power consumption is smaller than the household light bulb, and there is no tiny processor similar to the central processor.
To make silicon chips work like the brain, researchers are building non-digital chip systems that work as much as possible with real neural networks. A few years ago, Bochen designed a device called the Neurogrid, which was used to simulate the activity of 1 million neurons. The number of simulations was almost as many as the neurons in the bee's brain. To this day, "neuromorphic technology" has evolved over a quarter of a century, and it is not far from practical applications. This technology is expected to be used on any low-power and small-volume device, from smartphones to robots to artificial eyes and ears. In the past five years, such application prospects have attracted many researchers to participate in this field, and institutions in the United States and Europe have invested hundreds of millions of dollars in research funding.
Giacomo Indiveri of the Institute of Neuroinformatics (INI) believes that neuromorphic devices also provide powerful research tools for neuroscientists. In a real physical system, by observing which functions a neural model has or is missing, "scientists can understand why the brain structure is like this."
Bohen said that neuromorphic solutions should help break through the limitations of Moore's Law. For a long time, every two years or so, computer chip manufacturers need to double the number of transistors in a given space. The use of chip space has tended to the extreme. Soon, the circuit on the silicon chip will be too small and too tight to transmit a "pure" signal: electrons will "leak" from various components, causing silicon chips and The neurons are just as chaotic. Some researchers hope to solve this problem by using software patches, such as borrowing techniques like statistical error correction to make the Internet work smoothly. But in the end, Bohen said that the most effective solution still exists in our brains - something that has been around for millions of years.
Silicon cell
The idea of ​​neuromorphic was coined by Carver Mead of the California Institute of Technology in the 1980s. In terms of microchip design, Mead is a pioneer. He coined the term “nerve form†and was the first scholar to emphasize the brain's enormous advantages in energy efficiency.
Mead mimics the brain's low-power processing through "sub-threshold" silicon. At very small voltages, the normal chip cannot change the bit from "0" to "1", but the subthreshold silicon still has tiny, irregular electron currents flowing through the transistor, the fluctuation of this spontaneous current, its size and The variability is very similar to the flow of electrons formed by ions flowing in the neural circuit. Mead concluded that these currents can be controlled by microcapacitors, resistors, and other components that may form tiny circuits and exhibit the same electrical performance as real neurons. Chips can be connected into a decentralized grid, creating communication lines between components without going through a central processor, which works much like a real neural circuit in the brain.
In the 1990s, Mead and colleagues discovered that it is possible to build a network of silicon neurons (see "Inspiration from Biology"). The device receives an external current input signal through a junction that acts like a synapse in a real nervous system—a pulse that passes from one neuron to another through a synapse. Similar to real neurons, silicon neurons allow incoming signals to accumulate voltage inside the circuit. When the voltage reaches a certain threshold, the silicon neurons "discharge" and produce a series of "voltage spikes, which are instantaneous voltage peaks." These "voltage spikes" propagate along a wire. This wire acts like a neuron's axon (axon, which propagates along the axon). Although these spikes are "digital" and can only be in either the on or off state, silicon neurons operate like non-digital in the form of true neurons, so the current and voltage of silicon neurons It is not limited to a few discrete values, which is completely different from traditional chips.
The performance of silicon neurons reflects a key factor in brain energy conservation: Like the real brain, silicon neurons simply integrate the input signal before "discharging", which requires very little energy. In traditional computers, regardless of whether the chip is performing or not, it is necessary to continuously input energy to maintain the internal clock operation.
Mead's team also studied other neural circuits - the most striking is the silicon chip that mimics the retina of the eye. This device uses a detector array (50 x 50) to capture light. When the detector's activity is displayed on a computer screen, these "silicon cells" react to light, shadows, and motion in much the same way as neurons on the retina. Like the brain, this device only delivers meaningful information, saving energy: most cells in the retina react only when the light intensity changes. The advantage of this mode of operation is that it highlights the edge contours of moving objects to minimize the amount of data that needs to be transmitted and processed.
Coping with challenges
Bochen joined Mead's laboratory in 1990. In the early days, researchers were busy researching single-chip devices, such as the silicon retina. But by the end of the 1990s, "we wanted to build a 'brain', which required large-scale inter-chip information transfer." This is a huge challenge: the standard encoding algorithms for inter-chip communication are specifically designed to precisely coordinate digital signals. However, it cannot be applied to the more random spikes in the neuromorphic system. At the beginning of the 21st century, Bo Heng and other researchers invented circuits and algorithms that operate under this chaotic system, opening the way for a series of developments in large-scale neuromorphic systems.
In the first series of applications, one was a large simulator that provided a simple way for neuroscientists to study how the brain works. For example, in September 2006, Bochen launched the “Neural Grid†project, which hopes to simulate the activities of millions of neurons. Although it is only a small fraction of the human brain's 86 billion neurons, it is sufficient to simulate several closely related groups of neurons that are considered to be part of the computational unit in the human cerebral cortex. Bohen said that neuroscientists can construct any model of the cerebral cortex by coding the neural grid. Then, they will see that the brain model runs almost as fast as the brain – which is hundreds of times faster than traditional digital simulations. Researchers have used it to test some theoretical models of neurological function, such as working memory, decision making, and visual attention.
"Because it is highly similar to real brain neuron networks and has the same operational efficiency, Bohen's neural grid is far ahead of other large-scale neuromorphic systems." INI co-founder, one of the developers of silicon neurons Rodney Douglas said.
However, as Bohen himself said, no system is perfect. One of the biggest drawbacks of neural mesh is that synaptic connections are simple, with about 5,000 synapses per neuron, but synapses cannot be modified individually. This means that the system cannot be used to simulate a learning process, while in the brain, learning experiences modify the synaptic connections. Considering that the available space of the chip is limited, in a complicated circuit, if the working mode of each synapse is closer to the reality, the volume of the circuit component is required to be only one thousandth of the current, reaching the nanotechnology. range. This is currently not possible, but a new nanoscale storage device called memristors may solve this problem one day.
Another problem comes from the manufacturing process. In manufacturing, some small errors are unavoidable, which causes each neuromorphic chip to be slightly different at runtime. "These small errors are far less than what is observed in the brain," Bochen said, but it does mean that the neural grid program must allow substantial differences in the firing frequency of silicon neurons.
This problem has led some researchers to abandon Mead's original idea of ​​using "sub-threshold" silicon chips. They switched to more traditional digital systems. To a certain extent, these digital systems, by simulating the electrical properties of individual neurons, also have neuromorphic features, are more predictable and easier to program, but at the cost of higher energy consumption.
The SpiNNaker Project is a classic example. In 2005, Steve Furber, a computer engineer at the University of Manchester in the UK, launched the project. The system uses a very low power digital chip that can be found in many smartphones. The Spinnaker Program can currently simulate up to 5 million neurons. According to Fober, these neurons are simpler than the neurons in the neural grid and require more energy. The purpose of this system is similar to that of the neural grid: "Imitating the biological brain and running large brain models in real time" .
Another research and development direction is still to develop neuron chips, but it is trying to improve the speed of the chip. In a neural grid, neurons operate at the same speed as real neurons. The European BrainScaleS project, led by Karl Heinz Meier, a physicist at the University of Heidelberg, Germany, is developing a neuromorphic system that can now simulate 400,000 neurons and run faster than real brains. 10 000 times. This also means that in the same task, it consumes 10,000 times more energy than the brain, but the speed of the system is good news for some neuroscientists, Meyer said, "we can Simulate a day's nerve activity in 10 seconds."
Fobo and Meyer now have the funds to promote the development of the project, making the system bigger and more complete. They have all been selected into the European Union's Human Brain Project (HBP), which has been in operation for 10 years and has invested a total of 1 billion euros. Roughly, there will be 100 million euros for neuromorphological research, so the Fob team can expand the system and simulate 500 million neurons. At the same time, the Meyer team's goal is 4 million neurons.
But traditional methods can only do this. Since 2008, the US Defense Advanced Research Projects Agency has invested more than $100 million in the Synapse project to develop compact, low-power neuromorphic techniques. One of the project's main bearers, the cognitive computing team at the IBM Almaden Research Center, has used project funding to develop chips containing 256 digital neurons that can be used to build large systems.
Going to the application
Bohen is using his own method to achieve practical applications. I have to mention an unnamed project that he started in April 2013. The project is based on the Spaun model, a computer model of the brain that includes parts of the brain responsible for vision, movement, and decision making. Spaun relies on a neural circuit programming language developed 10 years ago by Chris Eliasmith, a theoretical neuroscientist at the University of Waterloo in Canada. The user only needs to specify the desired neurological functions, such as "generating a command to move the arm", and the system written by Elysses automatically designs the neural network to implement this function.
To verify that the system is working, Elysses and colleagues simulated the Spaun model on a traditional computer. They showed some simulation results, using 2.5 million simulated neurons, plus simulated retinas and hands, to complete the manual copying of numbers, recalling items in the list, calculating the next number in the given order, and other recognitions. Know the task. Bohen said that the ability of neural simulation has reached an unprecedented level. However, the Spaun simulation runs about 9 000 times slower than the actual running time of the brain, requiring 2.5 hours to simulate the brain's 1 second behavior.
Bohen introduced his proposal to Elysses: using real-time neuromorphic hardware to build a physical version of Spaun. “I am very excited,†Elysses said. In his opinion, this match is perfect. “You have peanut butter, and we have chocolate!â€
The US Office of Naval Research provided them with funding. Bohen and Elysses have formed a team that plans to build a small prototype within three years and build a comprehensive system within five years. .
Bo Heng said that they will use the neuromorphic retina and cochlea developed by INI for sensory input. For the output, they have a robotic arm. However, the construction of cognitive hardware will start from scratch. "This is not a new 'neural grid', but a completely new architecture." Because of the practicality, this architecture will not be highly similar to the real brain, but will rely on "very simple and very effective. Neurons, so that we can expand the number of neurons in the system to millions."
The system is designed for real-world applications. Bohen said, "Our vision is to build completely autonomous robots in a five-year period, they can interact with the environment in a meaningful way, complete tasks in real time, and only consume one mobile phone in terms of energy consumption. ". This device is more flexible, more adaptable, and consumes very little energy than today's automated robots.
Bohen added that in the longer term, his and Ellismis research will not only be used for robots, but also paving the way for the development of more compact, lower-power processors. The device will work with any computer system. If scientists have completely solved brain puzzles and found the key reason why the brain is so efficient, compact and powerful, this will be a big benefit for the computer industry. The invisible wall that blocks the development of chips will collapse and become smaller. The chip will be available.
Source: Global Science
Shaanxi Kang New Pharmaceutical co., Ltd. , https://www.bio-pharmacies.com