Researchers Achieve Neuronal Synaptic Plasticity and Neural Network Simulation Based on Inductor

The human brain is a highly interconnected, massively parallel and structurally complex network of ~ 10 ^ 11 neurons and 10 ^ 15 synapses. In neural networks, neurons are considered as computational engines of the brain, It accepts in parallel input signals from thousands of synapses connected to the dendrites, which are biological processes that produce changes in synaptic weights through specific patterns of synaptic activity, a process known as brain learning and Memory of the source.Simulation of synaptic plasticity and learning function, to build artificial neural network, is the future realization of neuromorphic brain computer key.In recent years, with the advent of new electronic devices and artificial intelligence technology, the use of a single electronic device The simulation of neuronal synaptic plasticity and learning functions has led to a new frontier research area - Synaptic Electronics.

Currently, synaptic electronics mainly uses the two-terminal resistance change device (memristor) and the three-terminal resistance change device (field-effect transistor) to simulate the function of the synapse.Its resistance and neuronal synaptic weight Touch connection intensity) has very similar changes and has been successfully used to simulate the synaptic plasticity and learning function.Recently, Institute of Physics, Chinese Academy of Sciences / Beijing Condensed State Physics Center National Laboratory of Sun Microsystems magnetic research group in The world first proposed a non-volatile circuit element based on the magneto-electric coupling effect (memtranstor) .This device is defined by the nonlinear relationship between charge and flux, the state value is electrically coupled That can be measured by measuring the magneto-electric coupling voltage value of the device.In the previous work, SUN Yang's research group has successfully demonstrated the two-state storage, polymorphic storage and Boolean logic functions at room temperature based on the memristor respectively. Compared with resistive devices, the memory device has the advantage of lower power consumption.

Recently, Sun Yang Research Group Shang Dashan, Ph.D. student Shen Jianxin and collaborator CAS academician, physics researcher Shen Baogen, Beijing Normal University Professor Wang Shouguo made new progress in the application of the memory device, the successful application of the memory device to the sudden In the field of electrochemistry, they achieved continuous and reversible changes of the coupling values ​​by adjusting the pulse trigger voltage and the number of pulses in a Ni / PMN-PT / Ni memory device with large magneto-electric coupling effect at room temperature. Contact weight increase and weakening behavior.Neurobiology studies show that learning function of neurosynxes following Hebrew, that changes in the weight of the synaptic depends on the activation of neurons before and after the connection of the synapses.They trigger waveforms by designing the pulse voltage, And superimposed the two sets of pulse waveforms to realize the synaptic plasticity behavior of pulse sequence-dependent plasticity (STDP) .On this basis, they built a 4 * 4 neural network based on the memory of a randomizer and used a random noise learning method to simulate The image static and dynamic learning function.The study for the first time in the world using the memristor to simulate the synaptic plasticity and learning function, The feasibility of constructing a low-power neural network based on a memory device is proved, which provides a completely new approach for the development of synaptic electronics and brain-like computing.

Relevant research results published in Advanced Materials.The research has been funded by the National Natural Science Foundation, Ministry of science and technology and Chinese Academy of Sciences.

Figure 1.a, the basic circuit components diagram; b, the characteristics of the coupling device; c, the working principle of the coupling; d, the magneto-electric coupling voltage (VME) with the magnetic field and polarization direction of the change.

Figure 2.a, schematic diagram of synapses; b, changes in the trigger potential of EPSP / IPSP with pulse triggering voltage; c, the number of pulses of EPSP / IPSP with pulse firing in the 2.5 kV / cm pulse voltage condition Variety.

Figure 3.a, b, pulse timing dependent plasticity and the corresponding pulse trigger waveform c, d, pulse timing dependent on the simplified form of plasticity and the corresponding pulse trigger waveform.

Figure 4.a, neural network diagram; b, trigger pulse waveform diagram; c, neural network learning before and after the distribution of synaptic weight; d, learning accuracy of the saturation value with the number of noise pixels changes.