Science paper: The human brain can perform XOR iso-dosing with just a single neuron.

New Zhiyuan report editor: Yuanzi [introduction to new Zhiyuan] the latest research by researchers from Humboldt University in Berlin and other institutions has confirmed that a new type of electrical signal has been found in the upper layer of human cortex. Each micro cell in the dendrite arm of cortical neurons can perform complex mathematical and logical operations, instead of the need for multilayer neural networks as previously thought.for example, a single neuron can easily solve the XOR operation that even the perceptron can't do."welfare: on January 16, Qian Qiao, senior research fellow of Tencent wechat artificial intelligence, taught you" recommendation system and data thinking ", and poked the link on the right side of" Xinzhiyuan we station open class "to learn more! Similar to the fine dendrites of plant roots, they radiate from the cell bodies of the cortical neurons in all directions.individual dendrites may process the signals they receive from adjacent neurons and then transmit them as input to the overall response of the cell.for a long time, we have studied the information processing ability of the brain by taking the trillions of brain neurons as a whole.but in recent years, more and more studies have begun to focus on the ability of individual neurons, and the results are surprising: these neurons seem to undertake much more computational responsibility than previously thought.the latest research by researchers from Humboldt University in Berlin and other institutions has confirmed that a new type of electrical signal has been found in the upper layer of human cortex. Each of the micro cells in the dendrite arm of cortical neurons can perform complex operations in mathematical logic, instead of needing multilayer neural networks as previously thought.for example, a single neuron can easily solve the XOR operation that even the perceptron can't do.the paper "democratic action potential and calculation in human layer 2 / 3 cortical nerves" published in science has attracted extensive attention. Konrad kording, a computational neuroscientist at the University of Pennsylvania, believes that it may prompt some computer scientists to reevaluate the strategies of artificial neural networks.the limitations of dumb neurons became popular in the 1940s and 1950s. Here's a picture of a dumb neuron as a simple integrator that sums inputs across the network.and the branches that extend from neurons (i.e., dendrites) receive thousands of signals from neighboring neurons.there are both excitatory and inhibitory signals.next, on the neuron body, all the signals are weighted and counted here. If the sum exceeds a certain threshold, the neuron will send out a series of electrical pulses (action potentials), which will directly stimulate the adjacent neurons.although it is not clear how much calculation the brain actually does in this way when processing information, it is this picture that makes scientists realize that a single neuron can also act as a logic gate, similar to a logic gate in a digital circuit! In theory, neural networks can perform any computation.of course, this kind of neuron model is limited.Bartlett Mel, a computational neuroscientist at the University of Southern California, said: "this is actually a point where neurons collapse into space.it does not have any internal moving joints. "at that time, scientists lacked experimental tools for recording from various components of a single nerve cell. the above model ignores that thousands of inputs into a given neuron enter the neuronal body along different dendrites, and the functions of these dendrites themselves may vary greatly, or more specifically, there may be some computational functions within these dendrites. fortunately, the situation has changed after 40 years. the new model of neuroscientist Christof Koch et al. Shows that a single neuron cannot express a single or unified voltage signal. on the contrary, the voltage signal decreases when it enters the neuron along the dendrite, and usually has no contribution to the final output of the cell. this separation of signals means that individual dendrites may process information independently of each other, which contradicts the previous neuron point hypothesis. this prompted Koch and other neuroscientists such as Gordon shepherd of Yale Medical School to model the structure of dendrites. in principle, neurons do not act as simple logic gates, but as complex multi unit processing systems. through a series of complex hypothesis mechanisms, they simulated how the tree tree carries a large number of logical operations. later, Mel and several colleagues were surprised to find that: dendrites can produce local spikes; dendrites have their own nonlinear input-output curve; dendrites have their own activation threshold (which is different from the overall threshold of neurons); dendrites themselves can act as gates or other units. Mel and his former graduate student yiota poirazi realized that this means that a single neuron can be thought of as a two-layer network! The dendrite will act as a nonlinear computing sub unit, collecting input and spitting out intermediate output. these signals will then bind in the cell, which will determine the way the whole neuron reacts. it is not clear whether the activity at the dendritic level actually affects the firing of neurons and the activity of adjacent neurons. in any case, the role of local processing in the whole neuronal system is beyond doubt. in terms of computational power, neurons are much stronger than we think. neuroscientist Shepherd said that most of the processing power in the cortex is actually below the threshold. in theory, almost any imaginable computation can be performed by a neuron with enough dendrites, each of which can perform its own nonlinear computation. and in the paper, scientists have taken a step forward, believing that not only the dendrites themselves, but also the micro cells in dendrites can perform complex calculations independently. unexpected spikes and Minsky's old confusion. Previous studies have focused on rodents, and Matthew larkum, a neuroscientist at Humboldt University, wants to know what's the difference between electrical signals in a large number of human neurons with longer dendrites? They took slices of brain tissue from layers 2 and 3 of the human cortex, which contained particularly large neurons and many dendrites. when they stimulated the dendrites with electricity, they found something strange. they saw unexpected, repetitive spikes. and these spikes seem to be completely different from other known neural signals. they are rapid and transient, like action potentials, caused by calcium flux. in comparison, our conventional action potentials are usually caused by sodium and potassium ions. moreover, although calcium induced signal transduction has been previously observed in rodent dendrites, these spikes can last longer. What's more, when more current is injected into dendrites, the discharge intensity of neurons is smaller? It's a very interesting phenomenon. in order to understand the possible impact of this new spike, the researchers constructed a model that could reflect the behavior of neurons. the model found that the dendritic crystal would have a peak signal in response to two separate inputs, but failed to do so when these input signals were combined. this is equivalent to a non-linear calculation called XOR (or XOR). If one input (but only one) is 1, then the binary output is 1. this discovery immediately resonated with the computer science community. for many years, it has been thought that XOR function is impossible in a single neuron! Computer scientists Marvin Minsky and Seymour paper, in their 1969 Book perceptrons, provide evidence that single-layer artificial networks cannot perform XOR. this conclusion is so shocking that many computer scientists attribute it to the continued downturn in neural network research before the 1980s. neural network researchers have finally found a way to avoid the difficulties proposed by Minsky and Papert, and neuroscientists have also found cases of these solutions in nature. poirazi has previously found that XOR may exist in a single neuron, which can be achieved by simply combining two dendrites. in this recent experiment, they even provided a reasonable biophysical mechanism for XOR execution in a single dendrite. of course, not all neurons in the processor are like this. according to Gidon, there are many smaller punctate neurons in other parts of the brain. so why does a single compartment within a neuron need to be able to do what an entire neuron or a very small network of neurons can do? The obvious possibility is that the neurons of multilayer neural network can have better processing ability and better learning and storage ability. poirazi thinks that perhaps there is a deep network in a single neuron, which is much more powerful in learning difficult problems or cognition. Such a strong single neuron may also help the brain save energy. kording added: maybe a neuron can calculate really complex functions. colding added:. according to poirazi, the Yong researchers believe that the above findings mark the need to rethink how to model the brain and its wider functions, and that it is not enough to focus only on the connectivity of different neurons and brain regions. the new results also seem to have an impact on problems in the field of machine learning and artificial intelligence. the artificial neural network relies on the point model, which regards the neurons as the nodes input and transfers the sum through the activity function. Gary Marcus, a professor of psychology and neuroscience at New York University, believes that the significance of this work is enormous, and the ability of a single neuron may exceed our imagination. may even reshape our view of "neural inspired" networks. he added: "the whole research rule of trying to get smart cognition from stupid neurons can be wrong. "reference link: