Circuit Implementations Of Memristor-based Neural Networks For Associative Memory And Pattern Recognition

With the development of machine learning,big data,and deep learning,traditional CMOS hardware platforms cannot satisfy the data processing requirements gradually due to the reaching end of Moore's Law and the limitation of Von-Neumann bottleneck.Hence,the hardware implementations of the next generation are explored to obtain good performances of data processing.In the emerging hardware platforms for data processing,memristor-based neural network circuits are considered as a promising candidate,improving data processing performances.Therefore,the studies on memristor-based neural network circuits are of great significance.It is the first step for the construction of a biology-like memristor-based hardware platform that applying memristor-based neural network circuits imitates biological features such as associative memory,STDP,homeostatic plasticity,etc.The existing memristor-based neural network circuits mimicking associative memory focus on imitating each procedure of associative memory while the changes of circuit parameters are unknown in the operation,which is disadvantageous for circuit extensions and applications.On the other hand,memristor-based neural network circuits are applied to accelerate computation,which are trained by algorithms to accomplish pattern recognition tasks.The existing memristor-based neural network circuits for pattern recognition have deficiencies that synaptic weights cannot adjust synchronously,the circuit structures do not match the corresponding training algorithms,etc.Based on the characteristics of associative memory and pattern recognition,this dissertation proposes the circuit implementations of memristor-based neural networks,improving the related deficiencies in the existing studies.The main research contents are presented as follows:First,a memristance changing circuit that consists of four MOS transistors and one memristor is proposed in the dissertation.The relation between the memristance change and the duration of the control signals is inferred according to the circuit structure.The memristance changing circuit is used as synaptic circuit to implement the memristor-based circuit mimicking full-function Pavlov associative memory.Next,a neuron circuit which is compatible with digital logic levels is designed by modifying the memristance changing circuit.By means of extending the neuron circuit with digital logic components and combining the characteristics of associative memory,the memristor-based neural network circuits are devised to realize the apple recognition and the recall of apple's features.For the proposed associative memory network circuits,the related parameter changes in the operation can be calculated approximately by the inferred equations,which is convenient for circuit extensions and applications.Then,this dissertation proposes a neuron circuit whose input signal is binary to achieve pattern recognition.Along with memristance variation,the corresponding synaptic circuit can represent negative weight,zero weight,positive weight,respectively.Meanwhile,the relation between the lasting time of control signal and the change of synaptic weight is obtained by the inference.Based on this neuron circuit,a single-layer character recognition network circuit and a multi-layer neural network circuit for pattern classification are designed in the dissertation.The character recognition network and the pattern classification network can achieve synchronous weight adjustment in an iteration.Namely,for an iteration,all the synaptic circuits that need to adjust can change weights at the same time,so the iteration is accomplished with one adjustment.This improves the training speeds of the memristor-based neural network circuits.Further,a neuron circuit whose input signal is continuous is proposed by modifying the synaptic circuit of the former binary one.According to this neuron circuit,a multi-layer memristor-based neural network circuit for pattern recognition is presented in the dissertation.By means of controlling the logic levels of the selecting passing signal,this multi-layer network can realize synchronous weight adjustment in an iteration.Then,an iris classification network circuit is proposed,based on the circuit structure of this multi-layer neural network.The circuit structure of the iris classification network matches back-propagation algorithm well,and it achieves a relatively high classification accuracy after training.Combining the characteristics of the associative memory and the proposed memristor-based neural network circuits,this dissertation proposes an associative-memory-based reconfigurable memristor-based neural network circuit.This neural network depends on the learning and forgetting of associative memory to cut off the connections of the corresponding synaptic circuits,realizing changeable circuit topology and the reconfiguration.A/D conversion is a necessary procedure in majority of the memristor-based neural network circuits for pattern recognition.Based on CMOS dual-slope A/D converter and the excellent properties of memristor,a memristor-based A/D converter is proposed in the dissertation.Compared with CMOS dual-slope A/D converter,the proposed one has compact circuit structure,relatively simple timing,and good integration on a chip.At the same time,the proposed A/D converter has good anti-interference performance and robustness.The summary of the dissertation is presented in the end.Meanwhile,this dissertation presents the studies which can be improved in the future.