Memory-based convolutional neural network system

A memory-based convolutional neural network system, comprising:

• an input module, a convolution layer circuit module composed of NOR FLASH crossbar as convolution kernels, a pooling layer circuit module based on NOR FLASH crossbar, an activation function module, a fully connected layer circuit module composed of NOR FLASH crossbar as synapses, a softmax function module and an output module, convolution kernel values or synapse weights are stored in a plurality of NOR FLASH units;
• the input module converts an input signal into a voltage signal required by the convolutional neural network, and transmits the voltage signal to the convolutional layer circuit module;
• the convolutional layer circuit module convolves the voltage signal with the convolution kernel values stored in the NOR FLASH units, and transmits the result to the activation function module;
• the activation function module activates the signal and transmits the result to the pooling layer circuit module;
• the pooling layer circuit module performs a pooling operation on the activated signal and the convolution kernel values stored in the NOR FLASH units, and transmits the result to the fully connected layer circuit module;
• the fully connected layer circuit module multiplies the pooled signal with the synapse weights stored in the NOR FLASH units to achieve classification, and transmits the classification result to the softmax function module;
• the softmax function module normalizes the output of the fully connected layer circuit module into a probability value, and then transmits the probability value to the output module as an output of the entire network,
• wherein the convolutional layer circuit module convolves the signal with the convolution kernel values stored in the NOR FLASH units, two NOR FLASH units are used for a convolution kernel in the circuit, the convolution kernel is converted into two matrices K+ and K−, an input signal X is correspondingly converted to a one-dimensional matrix, outputs of the K+ array are connected to a positive input of an operational amplifier included in the convolutional layer circuit module, outputs of the K− array are connected to a negative input of the operational amplifier, so that an output result is y=[(K+)−(K−)]*X, and an effective convolutional kernel value is (K+)−(K−), thereby achieving positive and negative convolution kernel values;
• in order to perform the convolution operation of the input signal in one step without an intermediate complex storage layer, when the convolution kernel values determined in the software are stored in the NOR FLASH units, the convolution kernel is expanded into a large sparse matrix that is able to be multiplied with the entire input signal;
• in the same row, gates of the NOR FLASH units are connected together to the input module interface, and sources of the NOR FLASH units are connected together to a driving voltage;
• in the same column, drains of the NOR FLASH units are connected together to the operational amplifiers, currents in the same column are brought together to achieve the addition calculation, and the outputs of the operational amplifiers are collected as a result of the convolution operation;
• the outputs of the operational amplifiers are transmitted to the activation function module.

A memory-based convolutional neural network system, comprising:

• an input module, a convolution layer circuit module composed of NOR FLASH crossbar as convolution kernels, a pooling layer circuit module based on NOR FLASH crossbar, an activation function module, a fully connected layer circuit module composed of NOR FLASH crossbar as synapses, a softmax function module and an output module, convolution kernel values or synapse weights are stored in a plurality of NOR FLASH units;
• the input module converts an input signal into a voltage signal required by the convolutional neural network, and transmits the voltage signal to the convolutional layer circuit module;
• the convolutional layer circuit module convolves the voltage signal with the convolution kernel values stored in the NOR FLASH units, and transmits the result to the activation function module;
• the activation function module activates the signal and transmits the result to the pooling layer circuit module;
• the pooling layer circuit module performs a pooling operation on the activated signal and the convolution kernel values stored in the NOR FLASH units, and transmits the result to the fully connected layer circuit module;
• the fully connected layer circuit module multiplies the pooled signal with the synapse weights stored in the NOR FLASH units to achieve classification, and transmits the classification result to the softmax function module;
• the softmax function module normalizes the output of the fully connected layer circuit module into a probability value, and then transmits the probability value to the output module as an output of the entire network,
• wherein the pooling layer circuit module mainly includes an average pooling operation and a maximum pooling operation, and has the same circuit structure as that of the convolutional layer circuit module, and convolution kernel values corresponding to these two operations are different;
• in the same row, gates of the NOR FLASH units are connected together to the activation function module, and sources of the NOR FLASH units are connected together to a driving voltage;
• in the same column, drains of the NOR FLASH units are connected together to an input of an operational amplifier included in the pooling layer circuit module, currents in the same column are brought together to achieve the addition calculation, and the results of the operational amplifiers are collected as a result of the convolution operation;
• the outputs of the operational amplifiers are transmitted to the fully connected layer circuit module.