CN109165730B - A State Quantization Network Implementation Method in Cross-Array Neuromorphic Hardware - Google Patents

Abstract

The invention belongs to the technical field of neural networks, and relates to a method for realizing a state quantification network in cross-array neuromorphic hardware. The method of the present invention is to quantify various parameters of the artificial neural network (weights, thresholds, leakage constants, setting voltage values, refractory period duration, synaptic delay duration and other parameters), and then quantify the quantized parameters. The state quantization network can be realized by mapping to the cross-array neuromorphic hardware, and then sending the pre-processed input data into the cross-array neuromorphic hardware. Through state quantization, the requirements of the cross-array neuromorphic hardware on the size of the storage unit, the number of storage levels, and the reliability are effectively reduced.

Description

A State Quantization Network Implementation Method in Cross-Array Neuromorphic Hardware

technical field

The invention belongs to the technical field of neural networks, and relates to a method for realizing a state quantification network in cross-array neuromorphic hardware.

Background technique

Neuromorphic computing is used to refer to brain-derived computers, devices, and models in stark contrast to the prevailing von Neumann computer architecture. This biomimetic approach creates highly connected synthetic neurons and synapses that can be used to model neuroscience theories and solve machine learning problems.

Neuromorphic circuit is one of the physical realizations of neural network models. It abstracts and simulates biological nervous systems at a high level by means of hardware, so as to achieve low power consumption, High adaptability and other characteristics.

Crossbars use memristors for data storage and parallel computing and as an important structure of neural network nodes. Crossbars are used to form large-scale integrated circuits. With a vertical crossover array, a large number of memristors can be placed together in parallel to form a memristor matrix. Under different voltage control, reading and changing the value of the memristor can obtain and set a weight matrix. Interleaved arrays are widely used in data storage and neural network learning.

In addition to the memristor, the unit at the intersection of the cross-array can also be composed of other devices, such as capacitors, transistors, variable resistors, etc., and can also form an array like a memristor for data storage or for cross-array neuromorphism. in hardware.

The prior art has at least the following problems:

In the currently implemented cross-array neuromorphic hardware, parameters such as synaptic weights and neuron parameters such as threshold, leakage constant, set voltage, refractory period, and synaptic delay require a lot of system storage resources. With the rapid expansion of circuit scale, this will inevitably become a major bottleneck for neuromorphic hardware in today's relatively scarce storage resources.

SUMMARY OF THE INVENTION

In view of the above problems, the present invention proposes a state quantization network implementation method in the cross-array neuromorphic hardware, which quantifies various parameters in the cross-array neuromorphic hardware, effectively reducing the size and storage capacity of the storage unit by the cross-array neuromorphic hardware. The requirements of series, reliability, etc. can strongly promote the application of cross-array neuromorphic hardware.

The technical scheme of the present invention is as follows:

S1: Select parameters and quantize them. Parameter quantization can be performed after the neural network training is completed, or it can be performed during the neural network training.

A: Quantize after neural network training is complete

Train artificial neural networks (including MLP, CNN, RNN, LSTM, etc.) under specific tasks and conditions to obtain parameters (including weights, thresholds, leakage constants, set voltage values, refractory period duration, synaptic delay duration) Wait);

The artificial neural network parameters obtained in S1 are quantized in the spiking neural network, and at least one parameter obtained by the above artificial neural network training is quantized in the spiking neural network, that is, several quantization states are used to replace all the parameters obtained by training. state. The parameters quantized in the spiking neural network are repeatedly adjusted so that the quantized spiking neural network achieves the predetermined function and performance, and the parameter quantization is completed.

B: Quantize during neural network training

When training the artificial neural network, the values of the parameters to be quantized (such as weights) are quantized, for example, the quantization values of the weights are set as -1, -0.4, 0, 0.4, 1, and then the artificial neural network is quantized. After training, the parameters are mapped to the corresponding spiking neural network, and the quantization parameters obtained from the training are adjusted or the quantization parameter values are re-selected for training, until the spiking neural network can achieve the predetermined function and performance, then the parameter quantization is completed.

S2: Mapping the quantized spiking neural network parameters in S1 into the cross-array neuromorphic hardware

The quantization parameters in the trained spiking network are mapped to the cross-array neuromorphic hardware, and different parameters are mapped to the corresponding control parts in the cross-array neuromorphic hardware. For example, the quantization weights are mapped to the cross arrays in the cross-array neuromorphic hardware; the quantization thresholds are mapped to the neuron threshold control part; the quantized leakage constants are mapped to the neuron leakage constant control parts; the quantized set voltage values are mapped to the neuron set voltage Control section; quantified refractory period duration maps to neuron refractory period duration control section; quantified synaptic delay duration maps to synaptic delay duration control section.

S3: Preprocess the input data, such as converting it into pulse input and coding, etc., and send the preprocessed input data to the cross-array neuromorphic hardware to realize the state quantization network.

Further, the cross-array in the above-mentioned cross-array neuromorphic hardware has multiple implementations;

Specifically, the unit at the intersection of the cross array can be realized by one transistor (N-type transistor, P-type transistor, floating gate transistor, synapse transistor, etc.);

Specifically, the unit at the intersection of the crossbar array can be realized by a semiconductor memory unit (such as a 6-tube SRAM unit);

Specifically, the unit at the intersection of the cross array can be realized by a capacitor;

Specifically, the unit at the intersection of the cross array can be realized by adding a selection transistor and a capacitor;

Specifically, the unit at the intersection of the crossed array can be realized by a memristor;

Specifically, the unit at the intersection of the crossover array can be realized by a selection transistor plus a variable resistor;

Specifically, the unit at the intersection of the crossbar array can be realized by a rectifier diode and a variable resistor.

The beneficial effect of the present invention is that, on the basis of converting the artificial neural network into an impulse neural network, the quantization of one or more parameters can be realized, and on this basis, the quantization parameters can be further mapped to the corresponding corresponding ones in the cross-array neuromorphic hardware. The control part is used to realize the state quantification network in the cross-array neuromorphic hardware, thereby realizing the state quantization at the hardware level, and effectively reducing the requirements of the cross-array neuromorphic hardware on the scale, storage level, reliability, etc. of the storage unit.

Description of drawings

1 is a flowchart of a method for implementing a state quantification network function in cross-array neuromorphic hardware provided by an embodiment of the present invention;

2 is a system structure diagram for realizing the state quantification network function in the cross-array neuromorphic hardware provided by an embodiment of the present invention;

Figure 3 is a cross-array form realized with capacitors;

Figure 4 is a cross-array form implemented with memristors;

Fig. 5 is the cross-array form realized with transistors and variable resistors;

FIG. 6 is a neuron model adopted in the embodiment of the present invention.

Detailed ways

In order to make the objectives, technical solutions and advantages of the present invention clearer, the embodiments of the present invention will be further described in detail below with reference to the accompanying drawings.

Referring to FIG. 1, an embodiment of the present invention provides a method for implementing a state quantization network in cross-array neuromorphic hardware, including the following steps:

S1: Select parameters and quantize them. Parameter quantization can be performed after the neural network training is completed, or it can be performed during the neural network training.

A: Quantize after neural network training is complete

The artificial neural network (including MLP, CNN, RNN, LSTM, etc.) is trained under specific tasks and conditions to obtain parameters, (including weights, thresholds, leakage constants, set voltage values, refractory period duration, synaptic delay duration, etc.);

The artificial neural network parameters obtained in S1 are quantized in the spiking neural network, and at least one parameter obtained by the above artificial neural network training is quantized in the spiking neural network, that is, several quantization states are used to replace all the parameters obtained by training. The quantized parameters in the spiking neural network are repeatedly adjusted so that the quantized spiking neural network can obtain the same function and almost the same performance as the original artificial neural network; then the parameter quantization is completed.

B: Quantize during neural network training

When training the artificial neural network, the values of the parameters to be quantized (such as weights) are quantized, for example, the quantization values of the weights are set as -1, -0.4, 0, 0.4, 1, and then the artificial neural network is quantized. Training, after the training is completed, the parameters are mapped to the corresponding spiking neural network, and the quantization parameters obtained from the training are adjusted until the spiking neural network can achieve the same function and almost the same performance as the original artificial neural network, then the parameter quantization is completed. .

S2: Mapping the quantized spiking neural network parameters in S1 into the cross-array neuromorphic hardware

Mapping the quantized parameters from the trained spiking network to the cross-array neuromorphic hardware, . Different parameters are mapped to different parts in the cross-array neuromorphic hardware. For example, the quantization weights are mapped to the cross-array in the cross-array neuromorphic hardware; the quantization threshold is mapped to the neuron threshold control part; the quantization leakage constant is mapped to the neuron The leakage constant control part; the quantized set voltage value is mapped to the neuron set voltage control part; the quantified refractory period duration is mapped to the neuron refractory period duration control part; the quantified synaptic delay duration is mapped to the synaptic delay duration control part.

S3: Preprocess the input data, convert the original input into pulse input and encode it, etc., and send the preprocessed input data to the cross-array neuromorphic hardware to realize the state quantization network.

The following takes weight quantization as an example for detailed description.

A corresponding artificial neural network is established for a specific task, and the neural network can be an artificial neural network of any model such as MLP, CNN, RNN, and LSTM.

Quantization after the neural network training is completed: Under certain conditions, the traditional artificial neural network training method is used to train it, and the parameters after training are obtained. Then map the parameters trained by the artificial neural network to the spiking neural network with the same topology structure, select the weight parameters that need to be quantified, and replace the previous ownership values with a limited number of weight states in the spiking neural network to realize the quantization of the weights. , use the quantized weights to test the spiking neural network, compare the performance of the spiking neural network and the corresponding artificial neural network at this time, if the spiking neural network reaches the performance index, the weight quantization ends, otherwise the weight quantization is re-taken , and test the spiking neural network again, and so on, until it reaches the performance target.

Quantization during neural network training: select several value states of the weights before training the artificial neural network, and train the artificial neural network. If the requirements can be met, the training is over; otherwise, the weights are changed. state, and retrain, and so on, until the artificial neural network meets the requirements and the training ends. Map the trained quantization weights and other parameters to the spiking neural network with the same topology as the artificial neural network, test the spiking neural network, and compare the performance of the spiking neural network and the corresponding artificial neural network at this time. If the network reaches the performance index, the weight quantization ends, otherwise the quantization weight value is adjusted, and the artificial neural network training and the mapping of the spiking neural network are performed again, and so on, until it reaches the performance index.

The trained spiking neural network is mapped to the corresponding weight control part in the cross-array neuromorphic hardware, that is, the cross-array unit. The relevant parameters of the cross-array unit are adjusted so that the weights in the cross-array neuromorphic hardware are the finite weight states obtained after the previous training of the spiking neural network.

The input data is preprocessed, converted into pulse input, encoded and sent to the cross-array neuromorphic hardware, and the weight quantization neural network function can be realized by the cross-array neuromorphic hardware.

FIG. 2 shows a system structure diagram of a state quantization network implemented by cross-array neuromorphic hardware in an embodiment of the present invention. As shown in the figure, the positive weights and negative weights are realized by their corresponding cross arrays respectively, and the input is divided into positive inputs +V _in,1 , +V _in,2 , +V _in,3 ,,,,,+V _in,n and negative input -V _in,1 ,-V _in,2 ,-V _in,3 ,,,,,-V _in,n (same sign as positive input). Among them, the original positive weight is retained in the positive weight cross array, and the negative weight is set to zero; the original negative weight is retained in the negative weight cross array, and the positive weight is zero, and the absolute value of the negative weight is taken. Each input generates an input with additional weights through the state of the corresponding unit of the crossover array (ie, the corresponding weight size), and a column of positive inputs and a corresponding column of negative inputs are respectively connected to the positive and negative inputs of the corresponding neurons. The relevant part of the internal sum-difference function of the element realizes the subtraction of the positive input and the negative input. Specifically, taking the input as positive input 1, positive input 2, negative input 9, and negative input 10 as an example, the corresponding input and output conditions of neuron 19 are analyzed, and the input and output conditions of other neurons can be similarly analyzed. Positive input 1 produces a corresponding weighted current input via 5 in the crossbar array; positive input 2 produces a corresponding weighted current input via 6 in the crossbar array; negative input 9 produces a corresponding weighted current input via 13 in the crossbar array Corresponding weighted current input; negative input 10 produces a corresponding weighted current input via 14 in the crossbar array. The column of the cross-array of positive weights corresponding to neuron 19 produces the total positive input 17, similarly the column of the cross-array of negative weights corresponding to neuron 19 produces the total negative input 18, and neuron 19 contributes to the total The positive input 17 is correlated with the total negative input 18 to realize the subtraction of the positive input and the negative input, and perform subsequent processing to finally generate the corresponding output.

The quantized weight of the spiking neural network is mapped to the weight control part of the cross-array neuromorphic hardware. Specifically, the quantized weight is mapped to the cross-array in the system structure shown in Figure 2. By adjusting the control part of each cross-array unit, Change the state of the cross array unit to realize weight quantization.

The following describes how different cross array units realize weight quantization

Figure 3 is a crossover array composed of capacitors. Specifically, each unit in the crossover array controls the connection strength between the input of the unit and the neuron, that is, the size of the weight. The capacitance unit in the crossover array and the amount of charge on the unit represent the size of the weight, and the current generated by the unit input is correspondingly larger. When the quantized weights of the spiking neural network are mapped to the cross-array neuromorphic hardware, the control unit of the capacitors in the cross-array is adjusted according to the value state of the quantization weights, so that the value state of the charge amount on the capacitor corresponds to the state of the quantization weight value one-to-one. , the weight quantization in the cross-array neuromorphic hardware can be realized by a cross-array consisting of capacitors.

Figure 4 is a cross-array composed of memristors. Specifically, the resistance state of the memristor represents the size of the weight, thereby controlling the magnitude of the current generated by the unit input. When the quantized weight of the spiking neural network is mapped to the cross-array neuromorphic hardware, the control unit of the memristor in the cross-array is adjusted according to the value state of the quantization weight, so that the resistance value state of the memristor is the same as that of the quantization weight value. One-to-one correspondence, the weight quantization in the cross-array neuromorphic hardware can be realized by a cross-array consisting of memristors.

Figure 5 is a cross-array consisting of transistors and variable resistors. Specifically, the resistance value of the variable resistor is changed to control the magnitude of the current generated by the unit input. When the quantized weights of the spiking neural network are mapped to the cross-array neuromorphic hardware, the control unit of the variable resistors in the cross-array is adjusted according to the value state of the quantization weights, so that the resistance value state of the variable electrical group is consistent with the quantization weights. If the states correspond one-to-one, the weight quantization in the cross-array neuromorphic hardware can be realized by a cross-array consisting of transistors and variable electrical groups.

FIG. 6 is a neuron structure model adopted in the embodiment of the present invention. The neuron implements positive and negative inputs while post-processing the final input. As shown in the figure, the neuron has positive input and negative input, and the positive input and negative input are realized by the weight processing part composed of resistors R ₁ , R ₂ , R ₃ , R ₄ , R ₅ , op amp 1 and op amp 2 The positive input is subtracted from the negative input to produce the final input. The final input is processed by subsequent units and produces the corresponding output. When the quantized spiking neural network parameters are mapped to the neuromorphic hardware, different parameters are mapped to the corresponding control parts of the neurons. The quantization threshold is mapped to the threshold control part of the neuron in the figure, that is, the adjustable voltage source part of the figure. The adjustable voltage source can generate multiple voltage values, corresponding to multiple thresholds of the neuron. According to the quantization threshold of the spiking neural network, the voltage generated by the adjustable voltage source is adjusted to change the threshold of the neuron, that is, the mapping of the quantization threshold is realized; the quantization leakage constant is mapped to R ₈ in the neuron in the figure, and by adjusting R ₈ The resistance value can change the leakage rate of the charge on the capacitor C ₂ , which can change the leakage constant of the neuron. When mapping the quantized leakage constant of the spiking neural network to the cross-neuromorphic hardware, adjust the resistor R ₈ accordingly according to the quantized leakage constant. The resistance value of the quantized leakage constant can be realized; the quantified refractory period duration is mapped to the refractory period duration control unit of the neuron in the figure, that is, the selection switch S in the figure. Selecting the switch S can adjust the conduction state and the holding time. During the holding time of the switch S in the discharging state, the neuron cannot respond to the external input, that is, the neuron is in the refractory period. The duration of the refractory period is quantified according to the spiking neural network. The holding time of switch S in each neuron firing can adjust the duration of the refractory period of the neuron, and the mapping of quantifying the duration of the refractory period can be realized.

Claims (7)

1. A method for realizing state quantification network in cross-array neuromorphic hardware, it is characterized in that, comprises the following steps: S1: select and quantify target parameters, the target parameters are one or more of weights, thresholds, leakage constants, setting voltage values, refractory period duration, and synaptic delay duration; parameter quantification is performed in the neural network After training is complete, or during neural network training, specifically: A: Quantize after the neural network training is completed, train the artificial neural network under the set target conditions to obtain the corresponding target parameters, map the target parameters to the spiking neural network and quantify the selected parameters, that is, use several The quantized state replaces all the states of the parameters obtained by training, and the quantized parameters in the spiking neural network are repeatedly adjusted so that the quantized spiking neural network achieves the predetermined function and performance, and the parameter quantization is completed; B: Perform quantization during neural network training, quantify the target parameters that need to be quantized, and quantify the values of the parameters to be quantized when training the artificial neural network, then train the artificial neural network, and map the parameters after the training is completed. into the corresponding spiking neural network, and adjust the quantization parameters obtained from the training or re-select the quantization parameter values for training, until the spiking neural network can achieve the predetermined function and performance, then the parameter quantization is completed; S2: Map the quantized spiking neural network parameters in S1 to the cross-array neuromorphic hardware: map the quantized parameters in the trained spiking network to the parameter control part corresponding to the cross-array neuromorphic hardware; The weights and negative weights are implemented by their corresponding cross arrays, respectively, and the input is divided into positive inputs +V _{in, 1} , +V _{in, 2} , +V _{in, 3} ,..., +V _{in, n} and negative inputs - V _{in, 1} , -V _{in, 2} , -V _{in, 3} , ..., -V _{in, n} , wherein, the positive weights are retained in the crossover array of positive weights, and the negative weights are set to zero; The original negative weights are retained in the crossover array, the positive weights are set to zero, and the absolute value of the negative weights is taken. Each input generates an input with additional weights through the state of the corresponding unit of the crossover array. A column of negative inputs is respectively connected to the positive input and negative input of the corresponding neuron, and the positive input and the negative input are subtracted through the interior of the neuron; S3: Preprocess the input data, and send the preprocessed input data to the cross-array neuromorphic hardware to realize a state quantization network; the preprocessing includes pulse input and coding. 2 . The method for implementing a state quantification network in cross-array neuromorphic hardware according to claim 1 , wherein a unit in the cross-array in the cross-array neuromorphic hardware is composed of a capacitor. 3 . 3 . The method for implementing a state quantification network in cross-array neuromorphic hardware according to claim 1 , wherein a unit in the cross-array in the cross-array neuromorphic hardware is composed of a memristor. 4 . 4 . The method for implementing a state quantification network in cross-array neuromorphic hardware according to claim 1 , wherein the unit in the cross-array in the cross-array neuromorphic hardware is composed of a selection transistor and a variable resistor. 5 . 5 . The method for implementing a state quantification network in cross-array neuromorphic hardware according to claim 1 , wherein a unit in the cross-array in the cross-array neuromorphic hardware is composed of one transistor. 6 . 6. The method for realizing a state quantification network in cross-array neuromorphic hardware as claimed in claim 1, wherein the unit in the cross-array in the cross-array neuromorphic hardware is composed of a selection transistor plus a capacitor at the cross-over unit of the cross-array neuromorphic hardware . 7 . The method for implementing a state quantification network in cross-array neuromorphic hardware according to claim 1 , wherein the unit in the cross-array in the cross-array neuromorphic hardware is composed of a rectifier diode and a variable resistor. 8 .

Images