Multi-task learning model based on spiking neural network, medium and equipment

Abstract

The invention relates to the technical field of artificial intelligence, in particular to a multi-task learning model based on a spiking neural network, a medium and equipment, and the method comprises the following steps: determining a target multi-task learning scene, and selecting a click rate prediction task, a skipping behavior prediction task, a collection behavior prediction task and a user scoring prediction task as a target task; obtaining an initial model of multi-task learning and initializing parameters, wherein the initial model comprises a pulse coding module, an expert layer module and a gating network module; constructing a data set; the initial model is trained by adopting the constructed data set, during training, data is transmitted and processed in a neural network in a pulse form, and model parameters are updated by using a back propagation algorithm in the training process. According to the multi-task learning model, SNN and MMOE are fused, and a data set commonly used for testing the learning ability of the model in multiple scenes with different relationships among tasks is constructed, so that the parameter quantity can be effectively reduced, the training time is shortened, and the learning ability of the model can be greatly improved.

Description

Technical Field

[0001] The present invention relates to the field of artificial intelligence technology, and in particular to a multi-task learning model, medium and device based on a pulse neural network. Background Art

[0002] With the continuous evolution of artificial intelligence technology, multi-task learning (MTL) has gradually become a key research direction to enhance the generalization ability of models. Among the many achievements of multi-task learning, the multi-gate mixture of experts (MMOE) model stands out with its unique gating mechanism. This mechanism can effectively achieve knowledge and information sharing between tasks, and performs well in maintaining task-specific representation capabilities.

[0003] However, most of the current research based on MMOE models is built on deep learning networks (DNNs). These DNN-based MMOE models have many problems in practical applications. On the one hand, they are mainly suitable for processing static data, and have limitations in fields such as time series prediction that require high data dynamics, and their performance is often unsatisfactory. On the other hand, the training cost of such models is extremely high, requiring a lot of computing resources and time investment. This is mainly due to its huge parameter group. Too many parameters not only increase the computational complexity of training, but also may cause problems such as overfitting, affecting the generalization ability of the model. Summary of the invention

[0004] The purpose of the present invention is to overcome the above problems or at least partially solve the above problems, and to propose a multi-task learning model, medium and device based on pulse neural network.

[0005] To achieve the above object, the first aspect of the present invention provides the following technical solution: a multi-task learning model based on a spiking neural network, comprising the following steps:

[0006] S1. Determine the target multi-task learning scenario and select the target tasks including click rate prediction, skip behavior prediction, collection behavior prediction and user rating prediction;

[0007] S2. Obtain an initial model for multi-task learning and initialize the parameters, wherein the initial model includes a pulse coding module, an expert layer module and a gated network module, the pulse coding module is used to convert input data into a pulse form, the expert layer module is used to extract and process features to provide feature representation for multiple tasks, and the gated network module is used to assign weights to each task to achieve personalized processing;

[0008] S3. Construct a data set with controllable relevance to the self-created task, which is used to perform performance testing on the initial model after obtaining and initializing parameters. The data set should include user data, song data, interaction data and context data, where the user data should include id, portrait, behavior and preference;

[0009] S4. Use the constructed data set to train the initial model. During training, data is transmitted and processed in the neural network in the form of pulses. The back propagation algorithm is used to update the model parameters during the training process.

[0010] In a preferred embodiment, in S2, when the initial model of multi-task learning is obtained and initialized, a gated network and a tower layer are set for each task, different weights are assigned to each task in a targeted manner, and a multi-expert layer is built using SNN (spiking neural network). The principle formula is expressed as follows:

[0011]

[0012] Among them, the gated network g k (x) = softmax(W gk x), whose input is the input feature and output is the weight on all Experts, x represents the input, W gk x∈R n×d represents a trainable matrix, n is the number of experts, d is the dimension of the feature, h k Indicates the tower layer.

[0013] In a preferred embodiment, in S4, data is transmitted and processed in the neural network in the form of pulses, and the processing method is as follows:

[0014]

[0015] s(t)=1,u(t)=u r2 ,ifu(t)≥u th

[0016] s(t)=0,ifu(t)≤u th

[0017] Where t represents the time step, τ is a constant, u and s represent the membrane potential and output peak, u r1 and u r2 are the resting potential and reset potential, respectively, j is the weight of the jth input synapse, is when the kth pulse of the jth input synapse is at T wThe moment when the state is 1 within this integral time window, K(·) is the kernel function representing the delay effect, T w is the integration time window, u th It is a threshold value, indicating whether to ignite once.

[0018] In a preferred embodiment, in S4, when the constructed data set is used to train the initial model, the STBP (space-time back propagation) training method is specifically used, and the principle formula is expressed as follows:

[0019]

[0020] In the formula, o represents pulse output, t represents time step, and n represents layer index. represents the delayed effect of membrane potential, f(·) is a step function, and this iterative LIF model contains all the behaviors in the original neuron model, including integration, triggering, and resetting.

[0021] The second invention of the present invention provides a multi-task learning model based on a spiking neural network. In S3, constructing a data set includes the following steps:

[0022] S3.1. Set the input feature dimension k and set two orthogonal unit vectors That is u1 T u2=0,||u1||2=1,||u2||2=1;

[0023] S3.2, set a proportional constant m and a correlation coefficient p, where -1≤p≤1, and generate two weight vectors ω1, ω2, where ω1=mu1,

[0024] S3.3. Randomly sample an input data point Each element is sampled from a standard normal distribution N(0,1);

[0025] S3.4. Generate two labels for the two regression tasks based on the weight vector:

[0026]

[0027] Among them, α i , β i , i=1,2…,n are parameters used to control the waveform shape in the sine function, ∈1,∈2~N(0,0.01) are independent and identically distributed noise terms. By adjusting the value of the correlation coefficient p between the weight vectors, the correlation between tasks is controlled. Repeat the operations of S3.3 and S3.4 until a sufficient data set is generated.

[0028] In a preferred embodiment, in S3.4, in order to indirectly control the correlation between tasks, cosine similarity is selected to be introduced to achieve:

[0029]

[0030] By changing the cosine similarity of w1 and w2, let cos(w_1, w_2) = p, where p is the preset cosine similarity value. In this way, the Pearson correlation between labels can be measured after generating data to evaluate the correlation between tasks.

[0031] The third aspect of the present invention provides an intelligent music recommendation system for user preferences, which adopts the above-mentioned multi-task learning model to predict the user's music preferences and can predict user preferences through user data, song data, interaction data and context data.

[0032] The fourth invention of the present invention provides a computer program, which includes computer program code. When the computer program code runs on an electronic device, the electronic device executes the above-mentioned multi-task learning model.

[0033] The fifth invention of the present invention provides a computer-readable storage medium having a computer program stored thereon, which can be executed by a processor to implement the steps of the above-mentioned multi-task learning model.

[0034] The sixth invention of the present invention provides an electronic device, comprising one or more processors and a memory, wherein the memory is used to store executable instructions, and one or more of the processors are configured to implement the steps of the above-mentioned multi-task learning model by executing the executable instructions.

[0035] Compared with the prior art, the present invention has the following beneficial effects: the multi-task learning model can effectively reduce the number of parameters and reduce the training time by fusing SNN with MMOE and constructing a data set that is commonly used to test the model's learning ability in multiple scenarios with different relationships between tasks, which can greatly improve the model's learning ability. BRIEF DESCRIPTION OF THE DRAWINGS

[0036] Figure 1 It is a flowchart of the present invention;

[0037] Figure 2 This is a schematic diagram of the model framework of the present invention;

[0038] Figure 3 Schematic diagram of the structure of a single neuron in the SNN of the present invention;

[0039] Figure 4Schematic diagram of the SNN structure inside the expert layer of the present invention. DETAILED DESCRIPTION

[0040] The following will be combined with the drawings in the embodiments of the present invention to clearly and completely describe the technical solutions in the embodiments of the present invention. Obviously, the described embodiments are only part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without creative work are within the scope of protection of the present invention.

[0041] See also Figure 1-4 The present invention provides a technical solution: a multi-task learning model based on a pulse neural network, comprising the following steps:

[0042] S1. Determine the target multi-task learning scenario and select the target tasks including click rate prediction, skip behavior prediction, collection behavior prediction and user rating prediction;

[0043] S2. Obtain an initial model for multi-task learning and initialize the parameters, wherein the initial model includes a pulse coding module, an expert layer module and a gated network module, the pulse coding module is used to convert input data into a pulse form, the expert layer module is used to extract and process features to provide feature representation for multiple tasks, and the gated network module is used to assign weights to each task to achieve personalized processing;

[0044] S3. Construct a data set with controllable relevance to the self-created task, which is used to perform performance testing on the initial model after obtaining and initializing parameters. The data set should include user data, song data, interaction data and context data, where the user data should include id, portrait, behavior and preference;

[0045] S4. Use the constructed data set to train the initial model. During training, data is transmitted and processed in the neural network in the form of pulses. The back propagation algorithm is used to update the model parameters during the training process.

[0046] Furthermore, the traditional MMOE network on which the multi-task learning model described in the present invention is based is a multi-task learning structure, in which the MoE used is a special neural network structure layer that combines the expert system and the integration concept. The model proposed in the present invention is composed of many "experts", each of which has a simple feedforward neural network and a trainable gating network, which selects a sparse combination of "experts" to process each input. It can automatically allocate parameters to capture information that can be shared by multiple tasks or information specific to a certain task without adding many new parameters for each task, and all parts of the network can be trained together through back propagation. The model principle can be expressed as:

[0047]

[0048] Among them, the input is x, E i represents the i-th expert network, G(x) i represents the i-th gating network, and the output y can be understood as the weighted sum of multiple expert opinions. MMOE uses multiple gating networks on this basis. The principle of the MMOE model can be expressed as:

[0049]

[0050] Among them, the gated network g k (x) = softmax(W gk x), its input is the input feature, and the output is the weight on all Experts. x represents the input, W gk x∈R n×d represents a trainable matrix, n is the number of experts, and d is the dimension of the feature. k Represents tower network. Among them, the expert layer is stacked by DNN, and the parameters of each layer are fully connected. This means that the more experts there are and the more complex the tasks are, the more the training parameters will explode, making it difficult to achieve the low power consumption and high performance that people expect from brain-like learning. Therefore, we focus our attention on the new generation of neural networks SNN, and try to use SNN to build an expert layer neural network on the basis of retaining the MMOE algorithm framework, so as to obtain a new generation of multi-task learning model. It is worth mentioning that the communication between pulse neurons is through binary events, not continuous activation values. S0 represents the pulses from the previous neuron, which are transmitted to the position of the dendrites through synapses, and are finally processed by the cell body (Soma). The specific processing method is:

[0051]

[0052] s(t)=1,u(t)=u r2 ,ifu(t)≥u th

[0053] s(t)=0,ifu(t)≤u th

[0054] Where t represents the time step, τ is a constant, u and s represent the membrane potential and output peak, u r1 and u r2 are the resting potential and reset potential, respectively, j is the weight of the jth input synapse, is when the kth pulse of the jth input synapse is at T wThe moment when the state is 1 within this integral time window, K(·) is the kernel function representing the delay effect, T w is the integration time window, u th It is a threshold value, indicating whether to ignite once.

[0055] SNN uses a sequence of pulses to transmit information, and each spiking neuron undergoes rich dynamic behaviors. Specifically, in addition to information propagation in the spatial domain, past history in the time domain also has a close impact on the current state. Therefore, compared with neural networks that mainly propagate through space and continuously activate, it usually has more temporal versatility. Since spike signals are only triggered when the membrane potential exceeds a threshold, the entire spike signal is usually sparse; in addition, since the spike value is binary, that is, 0 or 1, if the integration time window is adjusted to 1, the multiplication operation between the input and the weight can be eliminated; for the above reasons, compared with DNN networks with larger computational workloads, SNN networks can usually obtain lower memory and computational costs.

[0056] Furthermore, using the STBP (space-time back propagation) training method, the iterative version of the LIF model can be expressed as follows:

[0057]

[0058] In the formula, o represents pulse output, t represents time step, and n represents layer index. represents the delayed effect of membrane potential, f(·) is the step function, and this iterative LIF model contains all the behaviors of the original neuron model, including integration, firing, and resetting.

[0059] Furthermore, according to previous studies, the performance of multi-task learning depends largely on the correlation between tasks. However, it is very difficult to directly study the impact of the relationship between different tasks on the performance of multi-task learning. Therefore, we synthesize data to measure and control the correlation of tasks, so as to achieve the purpose of model comparison and testing. The steps are as follows:

[0060] (1) First, set the input feature dimension k and set two orthogonal unit vectors Right now

[0061] u1 T u2=0,||u1||2=1,||u2||2=1.

[0062] (2) Next, a proportionality constant m and a correlation coefficient p are set, where -1≤p≤1, and two weight vectors ω1, ω2 are generated.

[0063]

[0064] (3) Randomly sample an input data point where each element is sampled from a standard normal distribution N(0,1).

[0065] (4) Finally, two labels are generated for the two regression tasks based on the weight vector:

[0066]

[0067] Among them, α i , β i , i=1,2…,n are the parameters used to control the waveform shape in the sine function, ∈1,∈2~N(0,0.01) are independent and identically distributed noise terms.

[0068] By adjusting the value of the correlation coefficient p between the weight vectors, the correlation between tasks is controlled, and operations (3) and (4) are repeated until a sufficient data set is generated.

[0069] In order to indirectly control the correlation between tasks, we choose to introduce cosine similarity.

[0070]

[0071] By changing the cosine similarity of w1 and w2, let cos(w_1, w_2) = p, where p is the preset cosine similarity value. In this way, the Pearson correlation between labels can be measured after generating data to evaluate the correlation between tasks.

[0072] The Pearson correlation coefficient, usually expressed as r, is a statistical indicator that measures the degree of linear correlation between two variables. The value of the Pearson correlation coefficient ranges from -1 to +1, where +1 indicates a perfect positive correlation, -1 indicates a perfect negative correlation, and 0 indicates no linear correlation.

[0073] The calculation of the Pearson correlation coefficient is based on the covariance of two variables and their respective standard deviations.

[0074] The specific formula is:

[0075]

[0076] Estimating the sample covariance and standard deviation yields:

[0077]

[0078] Among them, X i and Y i represent the observed values ​​of two variables, and is their average. This formula captures the common tendency of two variables as they deviate from their means, thus providing a quantitative measure of their linear relationship.

[0079] In this study, we will use the Pearson correlation coefficient to analyze the correlation between two types of labels, and compare it with the cosine similarity of the weight vector to evaluate the correlation between tasks. The specific method is: for each cosine similarity value obtained by manual modulation, 10k data points with two labels y1 and y2 are generated, and the Pearson correlation between the two labels is calculated. After repeating this process 100 times, the obtained data is compared with the cosine similarity of the weight vector and an error graph is plotted.

[0080] In the linear case, that is, when the labels y1 and y2 are given by the linear equation

[0081] y1=w1 T x+∈1

[0082] y2=w2 T x+∈2

[0083] By definition, the Pearson correlation between labels is exactly equal to the cosine similarity p of the weight vectors.

[0084] In the nonlinear case, the labels y1 and y2 are defined by the nonlinear equations:

[0085]

[0086] At this point y1 and y2 are still positively correlated, although the correlation may not be exactly equal to the cosine similarity of the weight vectors.

[0087] In order to improve the performance of the model, solve the overfitting problem of the model and reduce the generalization error, we first perform hyperparameter tuning. We choose to use an automated tuner to complete hyperparameter tuning. The tuner uses an optimization algorithm based on Gaussian Process. This method effectively explores the best hyperparameter combination (such as learning rate, number of expert networks, and number of hidden units) by modeling functions in the hyperparameter space. At the same time, in order to effectively measure the performance of the model, we introduce AUC as a model evaluation indicator.

[0088] AUC, the full name of which is Area Under the Curve, is an indicator used to measure the performance of a binary classification model. Its value represents the model's ability to distinguish between random positive and negative sample pairs, ranging from 0 to 1. AUC is often used together with the ROC curve (Receiver Operating Characteristic Curve) to represent the area under the ROC curve.

[0089] The tuner continuously experiments with different hyperparameter combinations, uses the AUC score on the validation set to evaluate the effect of each training, and gradually approaches the hyperparameter setting that maximizes the AUC, thereby ensuring that the model performs best on the validation set to complete the hyperparameter tuning process.

[0090] At the same time, in order to achieve good classification performance on the validation set and test set, we tuned the model parameters and chose to implement it through the gradient descent algorithm. The trainable parameters of the model (such as the weights and biases of the neural network) are gradually updated according to the training data to minimize the loss function. During multiple trainings, parameter tuning optimizes the performance of the model on the training set and reduces the risk of the model falling into the local optimal solution through random initialization. By combining an efficient hyperparameter tuner and parameter tuning, the model can find the best global settings and weights.

[0091] According to an embodiment of the present invention, we provide a computer program product, which includes a computer program code. When the computer program code runs on an electronic device, the electronic device executes the above-mentioned multi-task learning model.

[0092] According to an embodiment of the present invention, we provide a computer-readable storage medium having a computer program stored thereon, wherein the computer program can be executed by a processor to implement the steps of the above-mentioned multi-task learning model.

[0093] According to an embodiment of the present invention, we provide an electronic device, comprising: one or more processors and a memory, wherein the memory is used to store executable instructions; the one or more processors are configured to implement the steps of the above-mentioned multi-task learning model by executing the executable instructions.

[0094] According to yet another embodiment of the present invention, we report and discuss experimental results on census-income data. We first apply the model to the UCI Census-Income dataset extracted from the 1994 Census database, which contains 299,285 examples of demographic information for US adults. There are 40 features in total, and we set these datasets as prediction targets, construct two multi-task learning problems, and calculate the absolute value of the Pearson correlation of the task labels over 10,000 random samples:

[0095] (1) Task 1: Predict whether the income exceeds $50,000;

[0096] Task 2: Predict whether the person’s marital status is never married;

[0097] Absolute Pearson correlation: 0.1768;

[0098] (2) Task 1: Predict whether his education level is at least college level;

[0099] Task 2: Predict whether the person’s marital status is never married;

[0100] Absolute Pearson correlation: 0.2373.

[0101] In this dataset, there are 199,523 training examples and 99,762 test examples. We further randomly split the test examples into a validation dataset and a test dataset in a 1:1 ratio. First, we removed the two features of education and marital status from the input features because they are regarded as labels in the following experiments. We then compared our model with the existing baseline methods. Since both sets of tasks are binary classification problems, we used the AUC score as the evaluation metric. Both groups of patients used the marital status task as an auxiliary task, the income task of the first group and the education task of the second group as the main task. In the hyperparameter tuning stage, we used the AUC of the main task on the validation set as the target. For each method, we used the hyperparameter tuner to conduct thousands of experiments to obtain the best hyperparameter settings, and finally trained the training dataset with random parameter initialization for 400 times, and reported the results on the test dataset.

[0102] Although embodiments of the present invention have been shown and described, it will be appreciated by those skilled in the art that various changes, modifications, substitutions and variations may be made to the embodiments without departing from the principles and spirit of the present invention, and that the scope of the present invention is defined by the appended claims and their equivalents.

Claims

1. A multi-task learning model based on spiking neural network, characterized in that: The following steps are involved: S1. Determine the target multi-task learning scenario and select the target tasks including click rate prediction, skip behavior prediction, collection behavior prediction and user rating prediction; S2. Obtain an initial model for multi-task learning and initialize the parameters, wherein the initial model includes a pulse coding module, an expert layer module and a gated network module, the pulse coding module is used to convert input data into a pulse form, the expert layer module is used to extract and process features to provide feature representation for multiple tasks, and the gated network module is used to assign weights to each task to achieve personalized processing; S3. Construct a data set with controllable relevance to the self-created task, which is used to perform performance testing on the initial model after obtaining and initializing parameters. The data set should include user data, song data, interaction data and context data, where the user data should include id, portrait, behavior and preference; S4. Use the constructed data set to train the initial model. During training, data is transmitted and processed in the neural network in the form of pulses. The back propagation algorithm is used to update the model parameters during the training process.

2. A multi-task learning model based on a spiking neural network according to claim 1, characterized in that: In S2, when the initial model of multi-task learning is obtained and initialized, a gated network and a tower layer are set for each task, different weights are assigned to each task in a targeted manner, and a multi-expert layer is built using SNN (spiking neural network). The principle formula is expressed as follows: Among them, the gated network g k (x) = softmax(W gk x), whose input is the input feature and output is the weight on all Experts, x represents the input, W hk x∈R n×d represents a trainable matrix, n is the number of experts, d is the dimension of the feature, h k Indicates the tower layer.

3. A multi-task learning model based on spiking neural network according to claim 2, characterized in that: In S4, data is transmitted and processed in the neural network in the form of pulses, and the processing method is as follows: s(t)=1,u(t)=u r2 ,ifu(t)≥u th s(t)=0,ifu(t)≤u th Where t represents the time step, τ is a constant, u and s represent the membrane potential and output peak, u r1 and u r2 are the resting potential and reset potential, respectively, w j is the weight of the jth input synapse, is when the kth pulse of the jth input synapse is at T w The moment when the state is 1 within this integral time window, K(·) is the kernel function representing the delay effect, T w is the integration time window, u th It is a threshold value, indicating whether to ignite once.

4. The multi-task learning model based on spiking neural network according to claim 3, characterized in that: In S4, when the constructed data set is used to train the initial model, the STBP (space-time back propagation) training method is specifically used, and the principle formula is expressed as: In the formula, o represents pulse output, t represents time step, and n represents layer index. represents the delayed effect of membrane potential, f(·) is a step function, and this iterative LIF model contains all the behaviors in the original neuron model, including integration, triggering, and resetting.

5. A multi-task learning model based on a spiking neural network according to any one of claims 1 to 4, characterized in that: In the S3, constructing the dataset includes the following steps: S3.1, set the input feature dimension k, and set two orthogonal unit vectors u1, That is u1 T u2=0,||u1||2=1,||u2||2=1; S3.2, set a proportional constant m and a correlation coefficient p, where -1≤p≤1, and generate two weight vectors ω1, ω2, where ω1=mu1, S3.

3. Randomly sample an input data point Each element is sampled from a standard normal distribution N(0,1); S3.

4. Generate two labels for the two regression tasks based on the weight vector: Among them, α i , β i , i=1,2…,n are parameters used to control the waveform shape in the sine function, ∈1,∈2~N(0,0.01) are independent and identically distributed noise terms. By adjusting the value of the correlation coefficient p between the weight vectors, the correlation between tasks is controlled. Repeat the operations of S3.3 and S3.4 until a sufficient data set is generated.

6. A multi-task learning model based on spiking neural network according to claim 5, characterized in that: In S3.4, in order to indirectly control the correlation between tasks, cosine similarity is introduced to achieve: By changing the cosine similarity of w1 and w2, let cos(w_1, w_2) = p, where p is the preset cosine similarity value. In this way, the Pearson correlation between labels can be measured after generating data to evaluate the correlation between tasks.

7. An intelligent music recommendation system based on user preferences, characterized by: The multi-task learning model as described in any one of claims 1 to 6 is used to predict the user's music preference, and the user preference can be predicted through user data, song data, interaction data and context data.

8. A computer program, characterized in that: The computer program includes computer program code, and when the computer program code runs on an electronic device, the electronic device executes the multi-task learning model as described in any one of claims 1 to 6.

9. A computer-readable storage medium, characterized in that: A computer program is stored thereon, and the computer program can be executed by a processor to implement the steps of the multi-task learning model described in any one of claims 1-6.

10. An electronic device, characterized in that: The method comprises one or more processors and a memory, wherein the memory is used to store executable instructions, and the one or more processors are configured to implement the steps of the multi-task learning model described in any one of claims 1 to 6 by executing the executable instructions.

Images