Memristive recurrent neural network circuit

Abstract

The invention provides a memristive recurrent neural network circuit, which comprises a neuron module circuit, a reverse summation circuit and a memristive cross synaptic array circuit, wherein a pre-synaptic neuron in the memristive cross synaptic array circuit transmits a current signal to the input end of the neuron module circuit through a memristor; and the output end of the neuron module circuit transmits a current signal to the reverse summing module circuit, and the reverse summing circuit reversely propagates the current signal to the negative electrode of the memristor cross synapse array circuit. The memristive neural network is combined with associative memory, and a full connection layer of the artificial neural network is realized based on memristive, so that the network area and power consumption can be greatly reduced.

Description

technical field [0001] The invention relates to the technical field of artificial intelligence, in particular to a memristive recursive neural network circuit. Background technique [0002] In recent years, a great deal of research has been done on the use of artificial neural networks (ANNs) to simulate the neural behavior of the human brain. On the one hand, many implementations of large-scale neuromorphic architectures have been proposed in order to find a model structure comparable to the computational power of the human brain. On the other hand, many studies are also trying to create intelligent agents with similar characteristics and behaviors of the human brain. [0003] Among them, the discrete Hopfield neural network is a structure of full feedback network. The characteristic of the full feedback network is that the input of any neuron is connected by the output of other neurons through a connection weight. Correspondingly, the output of any neuron is fed back to ...

AI Technical Summary

Problems solved by technology

[0006] The synapses in the above two Hopfield neural networks use resistors. Once the resistance value is fixed, the function of the entire network will be fixed accordingly. Since the weights of the synapses cannot be programmed flexibly, this kind of circuit will suffer greatly in application. At the same time, a large number of nonlinear devices are used in the discrete Hopfield neural network and the continuous Hopfield neural network, the power consumption of the entire circuit will be high and it is not conducive to the integration of the circuit

Images