Text pulse coding method and system based on spatial frequency characteristics

Abstract

The invention discloses a digital word pulse coding method (DWSFE) and a digital word pulse coding system fusing space and frequency characteristics, which are used for text information processing of a pulse neural network (SNN). The method comprises the following steps: constructing two vocabulary vector space modeling methods, namely interval semantic division vector coding (SSPE) and continuous semantic division vector coding (CSPE), based on a cerebral cortex semantic partitioning principle in neuroscience; converting the semantic vector into a pulse sequence with frequency characteristics through variable intensity Poisson coding (VIPE); and designing an unsupervised clustering SNN architecture containing a lateral suppression mechanism. According to the method, efficient pulse coding of text information is achieved, 100% classification accuracy is achieved under 400-dimensional word vectors, and compared with a traditional word embedding method, the application performance of SNN is greatly improved. The invention provides a low-power-consumption text processing system suitable for a neuromorphic computing device.

Description

Technical Field

[0001] The present invention belongs to the intersection of neuromorphic computing and natural language processing, and specifically relates to a text information pulse encoding method and system suitable for spiking neural networks. Background Art

[0002] Spiking neural networks (SNNs) transmit information through discrete pulse sequences, closely simulating the workings of biological neural systems. However, unlike traditional artificial neural networks that directly process continuous numerical values, SNNs can only accept discrete pulse signals in time and space. This makes efficiently encoding continuous semantic text information into pulse sequences a key challenge. Existing text encoding methods are primarily designed for traditional neural networks and have significant limitations when applied to SNNs.

[0003] The continuous vector representations generated by current mainstream word embedding technologies (such as Word2Vec, GloVe, etc.) cannot be directly used in SNNs and require additional pulse coding conversion. This conversion process not only increases the computational complexity, but may also lead to loss of semantic information. More importantly, these methods lack reference to the brain's language processing mechanism, making it difficult to achieve truly brain-like computing. Neuroscience research shows that the human brain's processing of semantic information has obvious spatial distribution characteristics, and different brain regions are specifically responsible for processing specific types of semantic information. This discovery provides important inspiration for designing more efficient text pulse coding methods.

[0004] Furthermore, traditional methods typically require a large number of training samples to achieve acceptable performance, which is inconsistent with the human characteristic of efficient language learning and also limits the application of SNNs in resource-constrained scenarios. Therefore, a new text pulse encoding method is urgently needed that can simultaneously meet the requirements of biological plausibility, computational efficiency, and semantic interpretability. This new method should fully draw on the research results of neuroscience and organically combine the spatial distribution characteristics of semantic information with the dynamic characteristics of pulse emission, thereby improving the efficiency and performance of SNNs in processing text information. Summary of the Invention

[0005] In order to solve the problem that continuous text semantic information in SNN is difficult to efficiently encode into spatiotemporal discrete pulse sequences, the present invention discloses a text pulse encoding method (Digit Words Spatial Frequency Encoding, DWSFE) that integrates spatial frequency features. Based on the spatial distribution law of brain semantics in neuroscience, a coding framework with both biological rationality and computational efficiency is constructed to enhance SNN's ability to deeply understand text semantics.

[0006] The specific plan is as follows:

[0007] A text information pulse encoding method suitable for a spiking neural network comprises the following steps:

[0008] S1. Semantic vector space modeling: Use the interval semantic partitioning vector encoding (SSPE) or continuous semantic partitioning vector encoding (CSPE) method to convert text words into sparse binary vectors;

[0009] In the present invention, a digital word set W is defined as shown below.

[0010] W={w0=“zero”,w1=“one”,...,w9=“nine”}

[0011] For each word w in the set W i ∈W,i∈[0,9], corresponding to a word pulse vector v of dimension D i , which is specifically defined as follows.

[0012] v i ∈{0,1} D ,i∈[0,9]

[0013] In order to enable the spiking neural network to learn the feature information contained in each word vector to the greatest extent, the vector v i The element value distribution in the word should reflect the differences in the part-of-speech characteristics between words. i Part of speech, this paper will pulse word vector v i Elements in different intervals are set to 1 to indicate activation, while elements in other positions are set to 0 to indicate inactivation. There are two main activation methods: Spaced Semantic Partitioning Vector Encoding (SSPE) and Contiguous Semantic Partitioning Vector Encoding (CSPE).

[0014] 1. Spacing Semantic Partitioning Vector Encoding (SSPE):

[0015] In the Semantic Spacing Vector Encoding (SSPE) method, the first activation bit is used as the core feature encoding unit to achieve semantic differentiation of digital words through the sparse binary encoding strategy with incremental offset. Specifically, each word w i The corresponding D-dimensional vector v i The following three encoding rules are met.

[0016] The first is the dynamic offset of the starting position. For each word w in the digital word set W i , the first activation position of its vector is shifted in increasing order according to the word order, that is, the vector vi The first activation position of is i·k0, where k0 is the basis interval. For example, when k0=1, the first activation position of v0 (corresponding to w0=“zero”) is the 0th dimension, the first activation position of v1 (corresponding to w1=“one”) is the 1st dimension, and so on.

[0017] The second is the periodic activation mode. After the first activation bit is determined, the subsequent activation bits are periodically expanded at a fixed interval k, where k is the semantic feature spacing coefficient, i.e., vector v i The non-zero element positions of are {i·k0,i·k0+k,i·k0+2k,...}, forming a sparse binary code with semantic discrimination in the vocabulary vector.

[0018] The third is feature decoupling. This encoding method ensures that the vectors of different numerical words are orthogonal in the feature space by distributing activation bits in increasing intervals, thereby enhancing the model's ability to capture discrete semantics.

[0019] The specific encoding method of SSPE is as follows.

[0020]

[0021] 2. Continuous Semantic Partitioning Vector Encoding (CSPE):

[0022] The CSPE method simulates the coordinated response of adjacent neuronal clusters to related semantic concepts by assigning continuous activation intervals to each word in the encoding vector. The encoding process of CSPE is divided into two stages.

[0023] The first stage is interval division, which evenly divides the D-dimensional vector into C continuous intervals, where C represents the number of semantic categories (in this experiment, the default setting is C = 10, corresponding to the 10 digital words from "zero" to "nine"). The length of each interval is

[0024] The second stage is position allocation, word w i The vector v i The activation area is limited to the i-th continuous interval, that is, the starting position s i =i·L, end position e i =(i+1)·L-1. The specific encoding formula of CSPE is as follows.

[0025]

[0026] S2. Frequency feature encoding: The semantic vector is converted into a time series signal with differentiated pulse emission frequency through the variable intensity Poisson encoding (VIPE) method;

[0027] Neuroscience research shows that the brain encodes information through neuronal firing rates. To simulate the brain's sensitivity to the gradient of neural activity intensity at the semantic level, this paper achieves differentiated settings of the word vector pulse firing frequency range by dynamically adjusting the Poisson coding intensity. i The element value is a binary feature of 0 or 1, and the Poisson encoding frequency is positively correlated with the intensity factor. In order to accurately control the discharge frequency, variable intensity Poisson encoding (VIPE) is used. In this method, each neuron in the input layer (i.e., VIPE neuron) uses a variable Poisson encoding intensity factor λ i , set differentiated strength intervals for vocabulary input vectors of different parts of speech.

[0028] The present invention is to use each word vector v i Poisson coding strength factor λ i The following uniform distribution interval model is set.

[0029] λ i ~U(λ base +Δλ·i-δ,λ base +Δλ·i+δ)

[0030] This model can effectively simulate the different words in neural activities and meet the biological interpretability requirements of neural information encoding. base represents the baseline discharge rate, and in this invention, the default value is set to 50. Δλ represents the average discharge rate gradient between categories, and in this invention, the default value is 15. δ represents the discharge rate fluctuation range within a category, and in this invention, the default value is 10.

[0031] For example, for the vector v0 of word w0 = "zero", its quantized Poisson coding strength factor is uniformly distributed between λ0 and U(40,60); for the vector v1 of word w1 = "one", its quantized Poisson coding strength factor is uniformly distributed between λ1 and U(55,75). Similarly, for the vector v9 of word w9 = "nine", its quantized Poisson coding strength factor is uniformly distributed between λ9 and U(175,195).

[0032] S3, Spiking Neural Network Processing: A two-layer spiking neural network (SNN) architecture with a lateral inhibition mechanism is used for unsupervised clustering learning;

[0033] The SNN clustering learning model in this invention adopts a two-layer biologically inspired network architecture, which includes an input semantic encoding layer and a feature learning output layer. This model realizes efficient feature extraction and unsupervised clustering learning of text pulse signals, as follows: 1. Semantic encoding layer

[0034] The core function of the semantic encoding layer is to convert words into pulse signals suitable for processing by spiking neural networks. This is accomplished in two steps. First, using SSPE or CSPE techniques, the input words are converted into semantic vectors containing spatial feature information. VIPE is then used to further encode the semantic vectors into time-series pulse signals with frequency characteristics. VIPE dynamically adjusts the Poisson encoding strength factor to assign a unique pulse frequency distribution to each word, simulating the firing rate gradient characteristics of neural activity and enhancing the biological interpretability and information expression capabilities of the encoding.

[0035] 2. Classification Learning Layer

[0036] The classification learning layer is responsible for clustering the pulse signals output by the semantic encoding layer to achieve semantic classification of vocabulary. Its key components and working mechanisms are as follows:

[0037] (1) Neuron model: The Leaky Integrate-and-Fire (LIF) model is selected as the basic neuron model, and an adaptive threshold mechanism is introduced. The LIF model retains the core biological characteristics of the HH model, simplifies the action potential process, improves computational efficiency, and adapts to large-scale simulation requirements. The mathematical model of the neuron discharge can be expressed as follows.

[0038]

[0039] Among them, V is the membrane potential, τ is the membrane potential time constant, which is larger in excitatory neurons than in inhibitory neurons. When the membrane potential of a neuron exceeds its membrane threshold V thres When the neuron fires a pulse, its membrane potential is reset to V rest .

[0040] In the adaptive threshold mechanism, the activation threshold of each neuron is composed of a basic threshold and a dynamic variable. When the neuron discharges, the dynamic variable accumulates in a pulse-dependent manner and is accompanied by an exponential decay as shown in the following formula, forming a negative feedback of the discharge frequency, achieving self-stabilizing regulation of the neuron's pulse firing rate, and ensuring stable operation of the network.

[0041]

[0042] where τ e and τ i are the time constants of the excitatory postsynaptic membrane potential and the inhibitory presynaptic membrane potential, respectively. Neurons regulate their membrane potential by controlling the intensity of input current or by varying the conductance of synaptic channels. This invention utilizes a conductance-driven synaptic transmission mechanism, in which the excitatory and inhibitory synaptic conductances exhibit time-varying properties. The following differential equation system describes the nonlinear coupling relationship between pulse events and the dynamic adjustment of conductance values.

[0043]

[0044] Among them E rest represents the resting membrane potential, E exc and E inh represent the equilibrium potentials of excitatory and inhibitory synapses, respectively, g e and g i are the conductances of excitatory and inhibitory synapses, respectively.

[0045] (2) Synaptic learning rule: Spike timing-dependent plasticity (STDP) is used as the unsupervised learning rule. To improve simulation efficiency, an online STDP learning paradigm based on weight association is established. A dual-channel activity tracking system is established for each synapse to monitor the spike events of pre- and post-synaptic neurons. When the pre-synaptic neuron triggers a spike, the synaptic weight decreases by a value proportional to the post-synaptic discharge activity; when the post-synaptic neuron discharges, the weight increases by a value proportional to the pre-synaptic discharge activity. After each input sample is trained, all synaptic weights are normalized to prevent the same neuron from learning different input features. The specific algorithm is the following equation.

[0046]

[0047] In the formula and are the time constants of the presynaptic and postsynaptic trajectories, respectively. post and η pre is the learning rate, w max and w min Represent the minimum and maximum values ​​of the synaptic weight allowed, A pre (t) and A post (t) are the presynaptic and postsynaptic trajectories, respectively.

[0048] (3) Lateral inhibition mechanism: A lateral inhibition mechanism based on neuron distance modulation is adopted to establish a mapping relationship between lateral inhibition strength and neuron Euclidean distance. When a neuron in the excitatory layer is activated, the inhibition strength on other neurons is determined according to the Euclidean distance between the two. In the initial stage of training, a smaller minimum lateral inhibition strength and a larger maximum lateral inhibition strength are set to encourage neurons with similar distances to learn similar features; as the training progresses, the lateral inhibition strength gradually increases, and eventually a global strong lateral inhibition state is formed to consolidate the features learned in the early stage. This mechanism avoids the defects of the traditional full inhibition strategy and improves the network training efficiency and feature modeling capabilities. Specifically, when neuron i in the excitatory layer is activated, its inhibition strength on neuron j is defined by the following formula.

[0049]

[0050] d max =max(d i,j )

[0051] Where i,j are used for neuron index; (x i ,y i ),(x j ,y j ) represents the coordinates of the neuron in the two-dimensional space of the excitatory layer, d i,j is the Euclidean distance between two neurons; d max represents the maximum Euclidean distance between neurons in the grid; C min 、C max Represent the minimum and maximum values ​​of the lateral inhibition weight respectively. By adjusting C min and C max The size of the lateral inhibition can be flexibly controlled.

[0052] (4) Network topology: The model adopts a classic two-layer unsupervised learning architecture, including a DWSFE input layer, an output layer, and a lateral inhibition layer. The DWSFE input layer converts English words into semantic vectors through SSPE or CSPE encoding, and then encodes them into time-series pulses by the VIPE neuron group. The output layer is composed of excitatory neurons, the number of which is variable; the lateral inhibition layer contains inhibitory neurons equal to the number of output layer neurons. The pulse signal generated by DWSFE encoding is fully connected to the excitatory neurons in the output layer, and the excitatory neurons in the output layer are bidirectionally connected to the inhibitory neurons in the lateral inhibition layer (one-to-one connection between excitatory neurons and inhibitory neurons, and full connection between inhibitory neurons and excitatory neurons, but excluding the excitatory neurons in the corresponding one-to-one connection position). When the membrane potential of an excitatory neuron in the output layer exceeds the threshold and emits a pulse, it activates its one-to-one connected inhibitory neuron. The inhibitory neuron exerts lateral inhibition based on Euclidean distance on other excitatory neurons in the output layer, prompting the output layer neurons to learn different or similar spatiotemporal features and realize semantic clustering.

[0053] A system for implementing a text information pulse encoding method applicable to a spiking neural network, comprising:

[0054] Semantic vector generation module: used to perform interval semantic segmentation vector encoding (SSPE) or continuous semantic segmentation vector encoding (CSPE) method;

[0055] Pulse conversion module: used to implement the variable intensity Poisson encoding VIPE method;

[0056] SNN processing module: an unsupervised learning network with adjustable lateral inhibition mechanism.

[0057] The beneficial effects of the present invention are:

[0058] The proposed text pulse encoding method (DWSFE) that integrates spatial frequency features achieves 100% classification accuracy with only 500 training samples under 400-dimensional word vectors through the synergistic effect of SSPE / CSPE spatial encoding and VIPE frequency encoding. Its performance and learning efficiency are significantly better than those of mainstream word embedding methods in the same SNN architecture. The biologically inspired SNN clustering model adopts a dynamic lateral inhibition mechanism and adaptive STDP learning, which enables the network to maintain more than 99% stable performance at a scale of 2500 neurons. This method combines biological rationality and computational efficiency, providing a low-power solution for text processing on neuromorphic chips. BRIEF DESCRIPTION OF THE DRAWINGS

[0059] Figure 1 Schematic diagram of the SSPE vocabulary vector encoding method proposed in this invention.

[0060] Figure 2 Schematic diagram of the CSPE vocabulary vector encoding proposed in this invention.

[0061] Figure 3 This is a schematic diagram of the VIPE pulse encoding method proposed in the present invention.

[0062] Figure 4 This is the structural diagram of the text clustering learning pulse neural network proposed in this invention.

[0063] Figure 5 Cluster distribution diagram of output layers of different sizes under 100-dimensional SSPE peak accuracy, where the number of neurons in the output layer is (a) 100, (b) 400, (c) 900, (d) 1225, (e) 1600, and (f) 2500.

[0064] Figure 6 The confusion matrix of the 100-dimensional SSPE network test set of various scales, where the number of neurons in the output layer is (a) 100, (b) 400, (c) 900, (d) 1225, (e) 1600, and (f) 2500.

[0065] Figure 7 Cluster distribution diagram of output layers of different sizes under 400-dimensional SSPE peak accuracy, where the number of neurons in the output layer is (a) 100, (b) 400, (c) 900, (d) 1225, (e) 1600, and (f) 2500.

[0066] Figure 8 The confusion matrix of the 400-dimensional SSPE network test set of various sizes, where the number of neurons in the output layer is (a) 100, (b) 400, (c) 900, (d) 1225, (e) 1600, and (f) 2500.

[0067] Figure 9 Cluster diagrams of output layers of various scales under CSPE training peak accuracy, including: (a) 100-dimensional word vector, 100-dimensional output layer, (b) 400-dimensional word vector, 100-dimensional output layer, (c) 100-dimensional word vector, 400-dimensional output layer, (d) 400-dimensional word vector, 400-dimensional output layer, (e) 100-dimensional word vector, 900-dimensional output layer, (f) 400-dimensional word vector, 900-dimensional output layer, (g) 100-dimensional word vector, 1225-dimensional output layer, (h) 400-dimensional word vector, 1225-dimensional output layer, (i) 400-dimensional word vector, 1600-dimensional output layer, (j) 400-dimensional word vector, 1600-dimensional output layer, (k) 400-dimensional word vector, 2500-dimensional output layer, (l) 400-dimensional word vector, 2500-dimensional output layer.

[0068] Figure 10 Confusion matrices for the 100- and 400-dimensional CSPE network test sets of various sizes, where the number of neurons in the output layer is (a) 100, (b) 400, (c) 900, (d) 1225, (e) 1600, and (f) 2500. The upper part of each sub-graph shows the 100-dimensional word vector and the lower part shows the 400-dimensional word vector. DETAILED DESCRIPTION

[0069] The present invention will be further explained below with reference to the accompanying drawings and specific embodiments. It should be understood that the following specific embodiments are only used to illustrate the present invention and are not used to limit the scope of the present invention.

[0070] This paper discloses a digital word pulse encoding method that integrates spatial and frequency features. The method primarily comprises the SSPE vocabulary vector encoding method, the CSPE vocabulary vector encoding method, the VIPE pulse encoding method, and a text clustering learning spiking neural network structure. The implementation of this method requires appropriate hardware support, including but not limited to computing devices capable of running deep learning algorithms, such as servers and high-performance computers, as well as storage devices for training data and model parameters.

[0071] The present invention uses the BindsNet framework developed based on Python language and PyTorch to carry out experiments. The experiment uses a training set containing 20,000 samples and a test set of 3,000 samples. The data set covers the ten basic categories of English numeric words from "zero" to "nine". In order to systematically evaluate the effect of semantic vector space modeling, the experiment compares the two vocabulary vector encoding methods of SSPE and CSPE. The word vector dimensions are 100 and 400 dimensions. The 100-dimensional space represents low-dimensional semantic representation, and the 400-dimensional space corresponds to high-dimensional semantic modeling. In SNN unsupervised clustering learning, the scale of the feature output layer has a significant impact on the model accuracy, among which the number of excitatory neurons directly determines the network's ability to extract spatiotemporal feature patterns from the input data. For this reason, in the experiment of the present invention, the number of excitatory neurons in the feature output layer is set to six scales of 100, 400, 900, 1225, 1600 and 2500 respectively to explore the impact of feature layers of different scales on model performance. The experimental process mainly includes the following steps:

[0072] Step 1: Dynamic pulse emission tracking and category labeling of neurons. Divide the training samples into batches of fixed size (default R = 250 samples per batch) for iterative training. After completing a batch of training cycles, the pulse emission frequency distribution of all excitatory neurons in the output layer when inputting samples of each category is counted. Classify each neuron into the sample category with the largest number of response pulses in the batch (i.e., the maximum response category). In subsequent training batches, continue to track the pulse emission pattern of neurons. If the classification result of a neuron in the current batch is consistent with the historical labeling, it is determined that it has formed a stable category representation ability; if not, it is considered that the neuron has not completed feature learning convergence.

[0073] Step 2: Design a classification accuracy evaluation metric. Calculate the proportion of stable neurons in the output layer to count the number of neurons with stable category representation capabilities. Calculate their ratio to the total number of neurons in the output layer as the classification accuracy evaluation metric for the current stage of the model. Accuracy monitoring: Record the classification accuracy after every 250 training samples to dynamically monitor the model's learning status.

[0074] Step 3: Early stopping based on the validation set. A validation set is created from the 20,000-sample word training set, with a 10% ratio (2,000 samples). After training for every 250 samples, the validation set is evaluated for classification accuracy. The validation set classification accuracy is recorded, and the network model parameters (including synaptic weights at each layer and the unsupervised clustering state of the output layer) with the highest validation set accuracy are stored. If no higher classification accuracy is achieved after five consecutive validation set evaluations, the training process is terminated and the testing phase begins.

[0075] Step 4: Model Testing and Generalization Performance Evaluation. After the early stopping mechanism is triggered, load the saved optimal network model parameters (i.e., the model with the highest validation set accuracy). Perform classification tests on 3,000 independent test set samples, calculate the model's classification accuracy, and evaluate its generalization performance.

[0076] Step 5: System overall operation process. In the encoding stage, the input text words are encoded by SSPE / CSPE to generate Figure 1 and Figure 2 The spatial semantic vector shown is then converted into Figure 3 The frequency characteristic pulse sequence shown in Figure 2. In the cluster learning stage, the pulse sequence input is as follows: Figure 4 The SNN network shown in the figure achieves unsupervised clustering learning through dynamic spike tracking and early stopping. During the performance evaluation phase, the model parameters are optimized using the validation set, and the classification accuracy and generalization ability of the model are finally verified on the test set.

[0077] Figure 8 and Figure 10 The experimental results of the confusion matrix show that both SSPE and CSPE encoding can achieve the highest classification accuracy of 100% in the 400-dimensional word vector space. Compared with the low-dimensional word vector ( Figure 6 SSPE-encoded 100-dimensional word embeddings achieved a classification accuracy of 86.6%. High-dimensional word embeddings provide richer semantic information, effectively improving the model's convergence speed and stability. The lateral inhibition mechanism enhances the network's ability to selectively learn features. When the network size is expanded to 2,500 output neurons, both encoding methods maintain over 99% stability, fully demonstrating the robustness of joint spatial-frequency encoding for large-scale networks.

[0078] By systematically comparing the training effects of different output layer sizes and word vector dimensions, the study found that the combination of 400-dimensional word vectors and medium-sized output layers (400-1600 neurons) has a significant difference in classification performance, clustering effect ( Figure 5 、 7 9) achieves an optimal balance between computational complexity and training efficiency. Furthermore, the introduction of an early stopping mechanism effectively avoids overfitting and ensures the model's generalization capability. Comparative experiments with traditional word embedding methods show that DWSFE achieves 100% test accuracy with only 500 samples for a 400-dimensional word vector. Its performance and learning efficiency significantly outperform mainstream word embedding methods such as Word2Vec, GloVe, and FastText on the same SNN architecture.

[0079] In summary, the present invention discloses a text pulse encoding method that integrates spatial frequency features, including spatial feature encoding based on SSPE / CSPE, VIPE frequency encoding, and unsupervised clustering learning of improved SNN. To address the problem that continuous text information is difficult to efficiently encode into discrete pulse sequences in spiking neural networks, the present invention achieves high-precision pulse representation and classification of digital words by simulating the semantic partitioning mechanism of the brain and the discharge characteristics of neurons. Experiments have shown that this method only requires 500 samples to achieve 100% test accuracy under a 400-dimensional word vector configuration, providing an innovative solution for text processing in the fields of intelligent voice interaction, neuromorphic computing, and so on.

[0080] It should be noted that the above content merely illustrates the technical idea of ​​the present invention and cannot be used to limit the scope of protection of the present invention. For ordinary technicians in this technical field, several improvements and modifications can be made without departing from the principles of the present invention. These improvements and modifications all fall within the scope of protection of the claims of the present invention.

Claims

1. A text pulse coding method integrating spatial frequency features, characterized in that: The following steps are involved: (1) Semantic vector space modeling: Use the interval semantic partitioning vector encoding (SSPE) or continuous semantic partitioning vector encoding (CSPE) method to convert text words into sparse binary vectors; (2) Frequency feature encoding: The semantic vector is converted into a time series signal with differentiated pulse emission frequency through the variable intensity Poisson encoding (VIPE) method; (3) Pulse neural network processing: A two-layer pulse neural network (SNN) architecture with a lateral inhibition mechanism is used for unsupervised clustering learning.

2. The method according to claim 1, characterized in that In the interval semantic segmentation vector encoding SSPE method, each word w i The corresponding D-dimensional vector v i The following three encoding rules are met: (1) Dynamic offset of the starting position; for each word w in the digital word set W i , the first activation position of its vector is shifted in increasing order according to the word order, that is, the vector v i The first activation position is i·k0, where k0 is the basis interval, v i ∈{0,1} D ,i∈[0,9]; (2) Periodic activation mode: After the first activation bit is determined, the subsequent activation bits are periodically extended at a fixed interval k, where k is the semantic feature spacing coefficient, i.e., vector v i The non-zero element positions of are {i·k0,i·k0+k,i·k0+2k,...}, forming a sparse binary code with semantic discrimination in the vocabulary vector; (3) Feature decoupling: By distributing activation bits at increasing intervals, we ensure that the vectors of different digital words are orthogonal in the feature space, thereby enhancing the model's ability to capture discrete semantics. The specific encoding method is as follows:

3. The method according to claim 2, characterized in that The Continuous Semantic Partitioning Vector Encoding (CSPE) method includes the following two stages: (1) Interval division stage: The D-dimensional vector is evenly divided into C continuous intervals, where C represents the number of semantic categories; the length of each interval is (2) Position allocation stage: word w i The vector v i The activation area is limited to the i-th continuous interval, that is, the starting position s i =i·L, end position e i =(i+1)·L-1; The specific encoding method is as follows:

4. The method according to claim 3, characterized in that In the variable intensity Poisson encoding VIPE method, each vector v i Poisson coding strength factor λ i It needs to obey a specific uniform distribution, namely: l i ~U(λ base +Dλ·i-δ,λ base +Dλ·i+d) Among them, the reference discharge rate λ base The value is 50, the average discharge rate gradient Δλ between categories is 15, and the discharge rate fluctuation δ within categories is 10.

5. The method according to claim 1, wherein The two-layer spiking neural network (SNN) architecture includes a semantic coding layer and a classification learning layer. The semantic coding layer is used to convert vocabulary into pulse signals suitable for spiking neural network processing. The classification learning layer is used to perform cluster analysis on the pulse signals output by the semantic coding layer to achieve semantic classification of vocabulary. Its components and working mechanism include: (1) A neuronal model based on the leaky integrate-trigger (LIF) model and a learning rule for spike-timing-dependent plasticity (STDP); (2) The network structure includes an input layer, an output layer, and a lateral inhibitory layer, wherein the excitatory neurons in the output layer are bidirectionally connected to the inhibitory neurons in the lateral inhibitory layer. The direction from excitatory neurons to inhibitory neurons is one-to-one connected, while the direction from inhibitory neurons to excitatory neurons is fully connected, excluding the excitatory neurons at the corresponding one-to-one connection positions; (3) The lateral inhibition strength is inversely proportional to the Euclidean distance between neurons. When neuron i in the excitatory layer is activated, its inhibition strength on neuron j is defined as follows: d max =max(d i,j ) Among them, (x i ,y i ),(x j ,y j ) represents the coordinates of the neuron in the two-dimensional space of the excitatory layer, d i,j is the Euclidean distance between two neurons; d max represents the maximum Euclidean distance between neurons in the grid; C min 、C max represent the minimum and maximum values ​​of the lateral inhibition weight, respectively.

6. A text pulse coding system for implementing the method according to any one of claims 1 to 5, characterized in that: include: Semantic vector generation module: configured to perform interval semantic segmentation vector encoding (SSPE) or continuous semantic segmentation vector encoding (CSPE) method; Pulse conversion module: configured to implement a variable intensity Poisson encoding VIPE method; Two-layer spiking neural network (SNN) processing module: an unsupervised learning network with an adjustable lateral inhibition mechanism.

Images