CN113989818B - Character classification method and system based on brain-like computing platform - Google Patents

Abstract

The invention relates to a character classification method and system based on a brain-like computing platform. The method comprises the steps of constructing a pulse neural network classification model; the impulse neural network model includes: the device comprises a coding layer, a maximum pooling layer and a classification layer; the maximum pooling layer comprises a plurality of pooled neurons and a maximum pooled output layer; the pooled neurons are connected pairwise, and the pooled neurons are connected with the pooled neurons in an inhibition manner; the pulse of the maximum pooling output layer is issued by a first pulse neuron; the maximum pooled output layer comprises a neuron; the neuron is connected with each pooled neuron; and (4) carrying out character classification on the MNIST handwritten digit data set by using a pulse neural network classification model. The invention can improve the precision of character classification.

Description

Character classification method and system based on brain-like computing platform

Technical Field

The invention relates to the field of character classification, in particular to a character classification method and system based on a brain-like computing platform.

Background

In the artificial neural network, the texture features can be extracted through maximum pooling, the influence of useless information is reduced, and relatively lower dimensionality is obtained by carrying out aggregate statistical processing on different features so as to avoid an overfitting phenomenon. I.e., reduce the number of parameters, reduce computational cost, and prevent overfitting. In convolutional neural networks, the maximal pooling has become an indispensable module in the network, and usually only the maximal value needs to be found in the pooled region, i.e. the maximal pooled output is considered. However, it is still a difficult point to realize the maximum pooling in the spiking neural network, because the output of the network is in a pulse form, it is difficult to calculate the maximum output, and in the spiking neural network, the maximum neuron is dynamically changed over time, which makes it difficult to compare the neurons with the neurons, and further makes it difficult for the brain-like hardware computing platform to compare the information between the neurons. In the classification task implemented based on the brain-like computing platform, the existing method cannot meet the characteristics that the brain-like hardware computing platform cannot compare, so that the method with the largest pooling cannot be applied to the brain-like hardware computing platform. In the actual classification task, the pooling layer is necessary, and average pooling is usually used instead of maximum pooling, and this alternative causes a loss of precision of the classification task and finally a reduction in classification precision.

Disclosure of Invention

The invention aims to provide a character classification method and a character classification system based on a brain-like computing platform, which can improve the precision of character classification.

In order to achieve the purpose, the invention provides the following scheme:

a character classification method based on a brain-like computing platform comprises the following steps:

constructing a pulse neural network classification model; the impulse neural network model includes: the device comprises a coding layer, a maximum pooling layer and a classification layer; the maximum pooling layer comprises a plurality of pooled neurons and a maximum pooled output layer; the pooled neurons are connected pairwise, and the pooled neurons are connected with the pooled neurons in an inhibition manner; the pulse of the maximum pooling output layer is issued by a first pulse neuron; the maximum pooled output layer comprises a neuron; the neuron is connected with each pooled neuron;

utilizing a pulse neural network classification model to classify characters of the MNIST handwritten digit data set; the MNIST handwritten digit data set includes pictures of handwritten digits 0-9.

Optionally, the step size of the maximum pooling is the same as the size of the maximum pooling area.

Optionally, the constructing a classification model of the spiking neural network further includes:

and training the weights of the impulse neural network classification model by using an ANN-to-SNN mode.

Optionally, the performing, by using a pulse neural network classification model, character classification on the MNIST handwritten digit data set specifically includes:

and converting the pixel values of the pictures in the MNIST handwritten digit data set into pulses in a Poisson coding mode.

A character classification system based on a brain-like computing platform, comprising:

the pulse neural network classification model building module is used for building a pulse neural network classification model; the impulse neural network model includes: the device comprises a coding layer, a maximum pooling layer and a classification layer; the maximum pooling layer comprises a plurality of pooled neurons and a maximum pooled output layer; the pooled neurons are connected pairwise, and the pooled neurons are connected with the pooled neurons in an inhibition manner; the pulse of the maximum pooling output layer is issued by a first pulse neuron; the maximum pooled output layer comprises a neuron; the neuron is connected with each pooled neuron;

the character classification module is used for classifying characters of the MNIST handwritten digit data set by using a pulse neural network classification model; the MNIST handwritten digit data set includes pictures of handwritten digits 0-9.

Optionally, the step size of the maximum pooling is the same as the size of the maximum pooling area.

Optionally, the method further comprises:

and the training module is used for training the weight of the pulse neural network classification model by using an ANN-to-SNN mode.

Optionally, the character classification module specifically includes:

and the coding unit is used for converting the pixel values of the pictures in the MNIST handwritten digit data set into pulses in a Poisson coding mode.

According to the specific embodiment provided by the invention, the invention discloses the following technical effects:

the invention provides a character classification method and a system based on a brain-like computing platform, wherein neurons in a pooling area in a maximum pooling layer in a constructed pulse neural network classification model are connected in pairs to realize a transverse inhibition function, after a neuron firstly sends a pulse, a pulse is sent to other neurons, and the pulse can inhibit the next pulse of other neurons, so that the pulse sent by other neurons can be inhibited, namely, only the firstly sent pulse is output in the range of sending a pulse by each neuron, the method that most of the maximum pooling needs to obtain the maximum information by comparison at present is avoided, the characteristics of a pulse neural network are combined, and unnecessary information is inhibited by the pulse sending mode, not only is pulse information for a single neuron preserved, it preserves the first pulse of multiple neurons over the largest pooled selection area. The method keeps the 'maximum' information carried by each neuron, is more consistent with the characteristics of the impulse neural network, effectively relieves a series of problems caused by the loss of the impulse information transmission of a single neuron on a hardware platform, and further ensures the classification precision in the whole handwritten digit classification network.

Drawings

In order to more clearly illustrate the embodiments of the present invention or the technical solutions in the prior art, the drawings needed to be used in the embodiments will be briefly described below, and it is obvious that the drawings in the following description are only some embodiments of the present invention, and it is obvious for those skilled in the art to obtain other drawings without inventive exercise.

FIG. 1 is a schematic flow chart of a character classification method based on a brain-like computing platform according to the present invention;

FIG. 2 is a diagram of a maximum pooling layer structure;

FIG. 3 is a schematic diagram of a classification model of a spiking neural network;

FIG. 4 is a diagram of a MNIST handwritten digit data set;

FIG. 5 is a diagram illustrating character classification results;

fig. 6 is a schematic structural diagram of a character classification system based on a brain-like computing platform according to the present invention.

Detailed Description

The technical solutions in the embodiments of the present invention will be clearly and completely described below with reference to the drawings in the embodiments of the present invention, and it is obvious that the described embodiments are only a part of the embodiments of the present invention, and not all of the embodiments. All other embodiments, which can be derived by a person skilled in the art from the embodiments given herein without making any creative effort, shall fall within the protection scope of the present invention.

The invention aims to provide a character classification method and a character classification system based on a brain-like computing platform, which can improve the precision of character classification.

In order to make the aforementioned objects, features and advantages of the present invention comprehensible, embodiments accompanied with figures are described in further detail below.

Fig. 1 is a schematic flow chart of a character classification method based on a brain-like computing platform provided by the present invention, and as shown in fig. 1, the character classification method based on the brain-like computing platform provided by the present invention includes:

s101, constructing a pulse neural network classification model as shown in figure 3; the impulse neural network model includes: the device comprises a coding layer, a maximum pooling layer and a classification layer; as shown in fig. 2, the maximally pooled layer comprises a plurality of pooled neurons and a maximally pooled output layer; the pooled neurons are connected pairwise, and the pooled neurons are connected with the pooled neurons in an inhibition manner; the pulse of the maximum pooling output layer is issued by a first pulse neuron; the maximum pooled output layer comprises a neuron; the neuron is connected with each pooled neuron; the step size of the maximum pooling is the same as the size of the maximum pooling area.

The specific implementation steps of the maximum pooling layer are as follows:

1. the first layer is pooled neurons, the size of the pooled region is determined, as shown in fig. 1, the size of the pooled region is selected to be 3, that is, the maximum pooling operation is performed on 3 neurons each time, and the step size of the maximum pooling is consistent with the size of the pooled region;

2. connecting the neurons in the pooling region with each other, wherein the neurons are in inhibition connection with other neurons, namely when any one neuron sends a pulse to the next layer, and simultaneously sends an inhibition pulse to other neurons in the pooling region, the lateral inhibition method can ensure that when a certain neuron in the pooling region sends a pulse first, other neurons receive an inhibition pulse, the inhibition pulse can inhibit the next pulse to be sent of other neurons, and the largest pooling output neuron only outputs the information of the first pulse neuron, so that the largest pooling operation can be avoided being completed by using a comparison method;

3. the second layer is a maximal pooled output layer, which includes a neuron, each pooled neuron of the first layer is connected to the maximal pooled output layer, the connection synapse weight between the neuron and other neurons is 1, and the threshold for the neuron to fire a pulse is set to 1 to ensure a smooth transfer of the pulse. When the first pulse neuron sends a pulse, the pooled output neuron outputs a pulse, and the inhibitory pulse sent by the first pulse neuron can inhibit the next pulse of other pooled neurons, so that when the neuron in the pooled region first sends a second pulse, the second first pulse is generated, a second round of lateral inhibitory pulses can be generated, and the pulses are transmitted regularly and continuously. Although the neurons of the first pulse are constantly changing, the pulses delivered to the largest pooled output layer are all first pulses.

S102, character classification is carried out on the MNIST handwritten digit data set by using a pulse neural network classification model; as shown in fig. 4, the MNIST handwritten digit data set includes pictures of handwritten digits 0-9.

The MNIST hand-written digital data set contains 6 ten thousand samples, and the test set contains 1 ten thousand samples.

After S101, the method further includes:

and training the weights of the impulse neural network classification model by using an ANN-to-SNN mode. The last layer of the network is a classification output layer, 10 neurons are provided in total, each neuron represents a classification category, and the final character classification result is obtained through the neurons; the input image and the classification result of the model obtained by visualizing the images in the dataset are shown in fig. 5.

S102 specifically comprises the following steps:

and converting the pixel values of the pictures in the MNIST handwritten digit data set into pulses in a Poisson coding mode.

Before the image is input into the network, the image needs to be coded at the second layer of the network. The size of the MNIST data set picture is 28 × 28, each pixel value corresponds to one neuron, the pixel values of the image are converted into pulses in a Poisson coding mode, and a neuron model used in the method is a Leaky integral-and-fire (LIF) neuron model. LIF neurons are already provided on our brain-like computing platform, so the LIF neuron model does not need to be rewritten in the classification model construction. LIF neurons and their membrane potentials

Is described by the following first order differential equation:

；

in the formula (I), the compound is shown in the specification,

the time constant of the membrane, referred to as the membrane time constant,

the sum of synaptic currents produced by the firing behavior of individual presynaptic neurons. When the membrane potential is

Greater than or equal to threshold potential

When activated, neurons produce excitations, i.e., fire pulses, which, along with conduction of action potentials, reset membrane potentials to

And remains unchanged during the absolute refractory period; when membrane potential

Below the threshold voltage, the neuron will not fire a pulse and the membrane potential will have a membrane-dependent time constant

Attenuation to know resting potential

。

Fig. 6 is a schematic structural diagram of a character classification system based on a brain-like computing platform provided by the present invention, and as shown in fig. 6, the character classification system based on the brain-like computing platform provided by the present invention includes:

the impulse neural network classification model building module 601 is used for building an impulse neural network classification model; the impulse neural network model includes: the device comprises a coding layer, a maximum pooling layer and a classification layer; the maximum pooling layer comprises a plurality of pooled neurons and a maximum pooled output layer; the pooled neurons are connected pairwise, and the pooled neurons are connected with the pooled neurons in an inhibition manner; the pulse of the maximum pooling output layer is issued by a first pulse neuron; the maximum pooled output layer comprises a neuron; the neuron is connected with each pooled neuron;

a character classification module 602, configured to perform character classification on the MNIST handwritten digit data set by using a pulse neural network classification model; the MNIST handwritten digit data set includes pictures of handwritten digits 0-9.

The step size of the maximum pooling is the same as the size of the maximum pooling area.

The invention provides a character classification system based on a brain-like computing platform, which further comprises:

and the training module is used for training the weight of the pulse neural network classification model by using an ANN-to-SNN mode.

The character classification module 602 specifically includes:

and the coding unit is used for converting the pixel values of the pictures in the MNIST handwritten digit data set into pulses in a Poisson coding mode.

The invention has two main advantages, one is that: in the classification task implemented based on a brain-like computing platform, the existing methods cannot meet the characteristics that a brain-like hardware computing platform cannot compare, so that the maximum pooling methods cannot be applied to character classification. In an actual classification task, a pooling layer is necessary, average pooling is usually used to replace maximum pooling, and the replacement mode can cause the loss of the precision of the classification task.

The advantages are two: the maximal pooling method of the present invention compared to existing methods not only preserves the "maximal" information of a single neuron, it preserves the information of multiple neurons. Currently, in the research on realizing the maximum pooling based on the impulse neural network, there are various schemes to define the "maximum" in the pooling window, such as selecting the neuron with the highest pulse-emitting frequency as the maximum; selecting the neuron with the maximum real-time pulse average release rate as the maximum; the neuron with the highest accumulated membrane voltage is selected as the maximum. Although these methods can obtain a good maximum pooling result on the brain-like simulation platform, on the brain-like hardware computing platform, information between layers is only transmitted through pulses, that is, the pulse emitting frequency and the number of emitted pulses of each neuron cannot be compared only by means of pulse transmission between layers, so that information carried by the "maximum" neuron cannot be output on the maximum pooling layer.

The maximum pooling scheme provided by the invention avoids most of the conventional methods for obtaining the maximum information through comparison, combines the characteristics of the impulse neural network, suppresses the unnecessary information through an impulse sending mode, and retains the maximum information. For most problems, the first-to-fire pulse on a neuron carries more characteristic information, so the scheme herein will retain this portion of the first-to-fire pulse for output to the next layer. Compared with other maximal pooling implementation schemes based on the impulse neural network, the method not only retains impulse information of a single neuron, but also retains the first impulse of a plurality of neurons on the maximal pooling selection area. The method is different from the traditional maximum pooling method which only outputs single maximum neuron information, further refines the concept of maximum and reserves the maximum information carried by each neuron, and the method better accords with the characteristics of a pulse neural network, effectively relieves a series of problems caused by the loss of the pulse information transmission of the single neuron on a hardware platform, and further ensures the classification precision in the whole handwritten digit classification network.

The embodiments in the present description are described in a progressive manner, each embodiment focuses on differences from other embodiments, and the same and similar parts among the embodiments are referred to each other. For the system disclosed by the embodiment, the description is relatively simple because the system corresponds to the method disclosed by the embodiment, and the relevant points can be referred to the method part for description.

The principle and the embodiment of the present invention are explained by applying specific examples, and the above description of the embodiments is only used to help understanding the method and the core idea of the present invention; meanwhile, for a person skilled in the art, according to the idea of the present invention, the specific embodiments and the application range may be changed. In view of the above, the present disclosure should not be construed as limiting the invention.

Claims (8)

1. A character classification method based on a brain-like computing platform is characterized by comprising the following steps:

constructing a pulse neural network classification model; the impulse neural network classification model comprises: the device comprises a coding layer, a maximum pooling layer and a classification layer; the encoding layer comprises a Poisson encoding operation for converting the pixel values of the picture into pulses; the neurons of the coding layer correspond to the pixel values of the pictures one by one; the maximum pooling layer comprises a plurality of pooled neurons and a maximum pooled output layer; the pooled neurons are connected pairwise, and the pooled neurons are connected with the pooled neurons in an inhibition manner; the pulse of the maximum pooling output layer is issued by a first pulse neuron; the maximal pooled output layer comprises a neuron; the neuron is connected with each pooled neuron; the classification layer has 10 neurons, each neuron represents a classification category, and the final character classification result is obtained through the neurons of the classification layer;

utilizing a pulse neural network classification model to classify characters of the MNIST handwritten digit data set; the MNIST handwritten digit data set comprises pictures of handwritten digits 0-9;

and inputting the input image into a pulse neural network classification model, and outputting a character classification result by the pulse neural network classification model.

2. The method for character classification based on the brain-like computing platform according to claim 1, wherein the step size of the maximum pooling is the same as the size of the maximum pooling area.

3. The character classification method based on the brain-like computing platform according to claim 1, wherein the constructing the impulse neural network classification model further comprises:

and training the weights of the impulse neural network classification model by using an ANN-to-SNN mode.

4. The method according to claim 1, wherein the character classification for the MNIST handwritten digit data set by using the impulse neural network classification model specifically comprises:

and converting the pixel values of the pictures in the MNIST handwritten digit data set into pulses in a Poisson coding mode.

5. A character classification system based on a brain-like computing platform, comprising:

the pulse neural network classification model building module is used for building a pulse neural network classification model; the impulse neural network classification model comprises: the device comprises a coding layer, a maximum pooling layer and a classification layer; the encoding layer comprises a Poisson encoding operation for converting the pixel values of the picture into pulses; the neurons of the coding layer correspond to the pixel values of the pictures one by one; the maximum pooling layer comprises a plurality of pooled neurons and a maximum pooled output layer; the pooled neurons are connected pairwise, and the pooled neurons are connected with the pooled neurons in an inhibition manner; the pulse of the maximum pooling output layer is issued by a first pulse neuron; the maximum pooled output layer comprises a neuron; the neuron is connected with each pooled neuron; the classification layer has 10 neurons, each neuron represents a classification category, and the final character classification result is obtained through the neurons of the classification layer;

the character classification module is used for classifying characters of the MNIST handwritten digit data set by using a pulse neural network classification model; the MNIST handwritten digit data set comprises pictures of handwritten digits 0-9;

and inputting the input image into a pulse neural network classification model, and outputting a character classification result by the pulse neural network classification model.

6. The system according to claim 5, wherein the maximum pooling step size is the same size as the maximum pooling area.

7. The system of claim 5, further comprising:

and the training module is used for training the weight of the pulse neural network classification model by using an ANN-to-SNN mode.

8. The system according to claim 5, wherein the character classification module specifically comprises:

and the coding unit is used for converting the pixel values of the pictures in the MNIST handwritten digit data set into pulses in a Poisson coding mode.

Images