Clustering representation self-supervised learning method based on hyperbolic flow homotopy embedding

Abstract

The invention discloses a clustering representation self-supervised learning method based on hyperbolic flow homotopy embedding, and belongs to the field of self-supervised learning, and the learning method specifically comprises the following steps: S1, constructing a self-supervised learning framework based on a prediction network and a target network, and taking ST-GCN as an encoder of the prediction network and the target network; s2, acquiring an original human body action skeleton sequence, performing sample enhancement on each group of acquired sequences through a behavior observation module, and performing feature representation on enhanced sample data; according to the method, the problem of solution space collapse can be effectively solved, that is, common solutions gathered near an original point cannot be learned, so that the features can be effectively distinguished in the solution space, the problem of early action overlapping existing in different actions is effectively solved, and the method has relatively good innovativeness in feature expression of self-supervised learning.

Description

Technical Field

[0001] The present invention relates to the field of self-supervised learning, and in particular to a clustering representation self-supervised learning method based on hyperbolic manifold homotopy embedding. Background Art

[0002] Three-dimensional human skeleton data has nonlinear high-dimensional spatiotemporal interaction information. At present, the self-supervised representation learning method based on skeleton often uses SimCLR as a contrastive learning paradigm to learn the representation of 3D skeleton in Euclidean space. However, Euclidean space is a special manifold in which the metric tensor of each point is a diagonal matrix and the elements on the diagonal are all 1. Therefore, mapping the data to the hyperbolic manifold space for representation learning can mine more spatiotemporal feature information. In addition, the observation rate unique to the human behavior prediction task makes it easy for self-supervised learning methods to learn trivial solutions and network collapse. This paper proposes an end-to-end one-stage online clustering representation contrastive learning method based on hyperbolic manifold. This paper adopts an asymmetric framework that discards negative samples and only learns from positive samples, avoiding the class conflict problem caused by the introduction of negative samples in traditional contrastive learning. This paper does not need to explicitly cluster a large number of data samples, but maps the model's representation of the data into a clustering matrix through a clustering head. The columns of the matrix are the probability that the data is divided into a certain class, avoiding the problem of being unable to learn online due to the reliance on large-scale samples and the defect of hard clustering being sensitive to outliers. After the clustering method is introduced, the entropy of the probability of input data being assigned to clusters is used as a constraint to avoid most samples being assigned to the same cluster, thereby eliminating the risk of model collapse caused by the framework of discarding negative samples. Summary of the invention

[0003] The purpose of the present invention is to solve the defects in the prior art and propose a clustering representation self-supervised learning method based on hyperbolic manifold homotopy embedding.

[0004] In order to achieve the above object, the present invention adopts the following technical solutions:

[0005] A cluster representation self-supervised learning method based on hyperbolic manifold homotopy embedding. The specific steps of this learning method are as follows:

[0006] S1. Build a self-supervised learning framework based on the prediction network and the target network, and use ST-GCN as the encoder of the prediction network and the target network;

[0007] S2, obtaining the original human action skeleton sequence, and performing sample enhancement on each group of sequences obtained through the behavior observation module, and then performing feature characterization on the enhanced sample data;

[0008] S3, construct a homotopy mapping function, learn the equivalence class of samples in space, and fit the continuous homotopy relationship between sample representations to construct embedded features;

[0009] S4, project the embedded features into the classification probability matrix through the clustering head, and use the classification probability as the soft label of the input data;

[0010] S5. Build a joint optimization network and define the final loss function to train and test the self-supervised learning framework.

[0011] Optionally, the specific steps of characterizing the enhanced sample data in S2 are as follows:

[0012] Q1.1: Get the original human action skeleton sequence S = {x1, x2, x3, ..., x T-2 ,x T-1 ,x T}, after random cutting and random rotation of each data enhancement and behavior observation module, a pair of enhanced sequences S is obtained q and S p , where x i ,i∈{1,2,3,…,T-2,T-1,T} represents the i-th frame skeleton in the entire sequence, and T represents the number of frames in the skeleton sequence;

[0013] Q1.2: The prediction network and the target network use ST-GCN as encoders respectively, where the prediction network encoder f θ For prediction branch, the target network encoder f ξ For the target branch, the sequence S will be enhanced later q and S p Input the prediction network encoder f respectively θ and the target network encoder f ξ and through the formula y θ =f θ (s q ;θ) and y ξ =f ξ (s p ξ) obtains the feature vector after characterization, and then uses the momentum update method to update the target network encoder f ξ The specific update formula of the parameter ξ is as follows:

[0014] ξ=α·ξ+(1+α)·θ

[0015] Where α represents the momentum factor, and α∈[0,1];

[0016] Q1.3: Update the mapping layer of the prediction network through momentum to obtain the mapping layer of the target network. The momentum update method is consistent with the target network encoder. After obtaining the feature vector, use the formula z θ =g θ (y θ ) and zξ =g ξ (y ξ ), projecting the obtained feature representation to a projection layer of lower dimension through two mapping layers, where the two projection layers have exactly the same structure, both consisting of a fully connected layer, a batch normalization layer, and a ReLU nonlinear activation layer, where the dimension of each layer feature is 128;

[0017] Q1.4: The output z of the prediction network encoder is θ Input prediction layer q, prediction layer q passes through o θ =q(z θ ) to obtain the output result of the prediction network, and then θ With z ξ Normalize and calculate the mean square error between the two, then optimize the mean square error between the two to bring the two features closer in the feature space to obtain the corresponding loss function L basic , its specific manifestations are as follows:

[0018]

[0019] Among them, L basic The input is obtained later by exchanging data through two network branches.

[0020] Optionally, the specific steps of constructing the homotopy mapping function in S3 are as follows:

[0021] Q2.1: Assume two topological spaces X and Y, and f and g are continuous mappings of X and Y respectively. If there exists a continuous mapping H:X×I→Y such that H(x,0)=f(x), H(x,1)=g(x), then H is used as a homotopic mapping that transforms f into g, where I=[0,1]. Then the human motion skeleton sequence is described as a time series set of joint 3D positions, and each frame describes an action posture by coordinates;

[0022] S2.2: Based on the constraints implied between the coordinates, the set of all frame data is placed in a low-dimensional manifold in a high-dimensional space, and each sample is represented as a discrete curve in space. The input of the prediction network and the target network is assumed to be a continuous function defined in the manifold space in time, and this continuous function is passed through the prediction network encoder f θ and the target network encoder f ξ Projected into a point in the manifold space, and then using the projection function g θ and ξ Projected into the same space;

[0023] S2.3: Define p∈Sn as a hyperbolic embedding variable, ||p||∈[0,1), construct the homotopy function of the hyperbolic space, and establish a relationship between the hyperbolic space and the projection function. The specific calculation formula of the homotopy function is as follows:

[0024] Hy(x;p)=||p||g θ +(1-||p||)g ξ

[0025] In the formula, when ||p||=0, Hy(x;0)=g ξ ;and When ||p|| changes from 0 to 1, the function Hy(x; p) changes from g ξ Continuously change to g θ , this continuous change is called homotopy, expressed as Hy(x; p):g θ :g ξ .

[0026] Optionally, the specific steps for constructing the embedding features described in Q3 are as follows:

[0027] S3.1: The n-dimensional Poincare model in negative curvature space is selected as the target embedding space, where the Poincare model is specifically expressed as:

[0028]

[0029] For a pair of positive samples after data enhancement, the prediction network encoder f θ and the target network encoder f ξ The encoded feature vector z θ and z ξ , encode the two eigenvectors and obtain h after homotopy mapping θ and h ξ As the vector to be embedded;

[0030] S3.2: The vector to be embedded is normalized by L2 to obtain as well as Two sets of unit vectors are then used to calculate their similarity using the inner product form and used as the norm. as well as It is embedded in the Poincare model, where λ∈(0,1), and the parameterized Poincare model embedding is defined as v=λr, located in S n middle;

[0031] S3.3: Definitions Be the Mbius sum on the Poincare model, and let η,ξ∈S n , then the summation formula is defined as:

[0032]

[0033] Based on the defined summation formula, it can be seen that the operation does not satisfy the commutative law. Since the Poincare model is a conformal model with the property of preserving angles, based on the summation formula 1+2η,ξ>+|ξ 2 and 1-η 2 From the two coefficients, we can see that if the norms of η and ξ are closer to 1, the summation result will be closer to η, and vice versa.

[0034] Optionally, the specific steps of projecting the embedded features into the classification probability matrix through the clustering head and using the classification probability as the soft label of the input data in S4 are as follows:

[0035] S4.1: The clustering head is composed of a fully connected layer, a batch normalization layer, and a ReLU nonlinear activation layer. The dimension of the internal features of each layer is 128. The constructed clustering head is used as a multi-layer perceptron. The output data decibels of the prediction network and the target network are clustered through the multi-layer perceptron to obtain the clustering matrix Y a and Y b ;

[0036] S4.2: Use the cosine distance to measure the similarity between cluster pairs. Consider the same cluster as a positive sample pair and different clusters as negative samples, and use them as soft labels for the input data. The specific calculation formula of the cosine distance is as follows:

[0037]

[0038] Among them, (i,j)∈[1,M], and then the loss function is used To pull away clusters With the exception of The distance of all clusters other than C is the temperature parameter of clustering, and finally the loss function at the clustering level is obtained as follows:

[0039]

[0040] Among them, H(Y) represents the entropy of clustering assignment probability. Adding assignment entropy can effectively prevent most clusters from being assigned to the same class.

[0041] Optionally, the specific steps of constructing the joint optimization network described in S5 and defining the final loss function to train and test the self-supervised learning framework are as follows:

[0042] S5.1: After the self-supervised learning framework based on a prediction network that updates parameters through back-propagation and a target network with gradient cut-off is built, the output of the target network is used as the prediction network target, and after each round of gradient update, the new parameters are updated by exponential sliding average to update the parameters of the target network;

[0043] S5.2: Pass the prediction network encoder f θ and the target network encoder f ξ The two sets of feature vectors obtained are passed through the projection layer g θ (·) and g ξ (·) is transformed to the same space, and then the prediction network is connected to a multi-layer perceptron q θ (·) Make the output close to the target network, and obtain the contrastive learning loss through MSE loss. The specific calculation formula is as follows:

[0044]

[0045] Among them, z θ =g θ (f θ (x θ )),z ξ =g ξ (f ξ (x ξ ));

[0046] S5.3: Learn the homotopy transformation through a multilayer perceptron h(·) and use the similarity of the two representations as the norm The features are projected onto the Poincare model, and then the Mbius sum of the Poincare model is used to represent the representative points of the sequence in the hyperbolic space:

[0047]

[0048] in, It is obtained by exchanging the input of the prediction network and the target network. Its calculation process is consistent with λ, so the corresponding loss function is obtained as follows:

[0049]

[0050] L Hyp Combined with the original loss function, the final loss function is defined as:

[0051]

[0052] Compared with the prior art, the present invention has the following beneficial effects:

[0053] This clustering representation self-supervised learning method based on hyperbolic manifold homotopy embedding uses hyperbolic manifold as the representation space of features, so that the features present a tree structure in the representation space, characterize the hierarchical relationship of samples, and use the Mbius and method to further improve the target encoding. At the same time, by establishing the connection between the characteristics of the hyperbolic space and the homotopy trajectory, the samples can be adaptively homotopically classified according to the differences in the output of the samples under different mappings, which can effectively solve the problem of solution space collapse, that is, no trivial solutions clustered near the origin will be learned, so that the features can be effectively distinguished in the solution space, effectively solving the problem of early action overlap of different actions, and has good innovation in the feature expression of self-supervised learning. BRIEF DESCRIPTION OF THE DRAWINGS

[0054] The accompanying drawings are used to provide further understanding of the present invention and constitute a part of the specification. They are used to explain the present invention together with the embodiments of the present invention and do not constitute a limitation of the present invention.

[0055] Figure 1 A flowchart of the clustering representation self-supervised learning method based on hyperbolic manifold homotopy embedding proposed by the present invention;

[0056] Figure 2 This is the technical roadmap for the clustering representation self-supervised learning method based on hyperbolic manifold homotopy embedding proposed in this invention. DETAILED DESCRIPTION

[0057] Reference Figure 1-2 , a cluster representation self-supervised learning method based on hyperbolic manifold homotopy embedding, the specific steps of this learning method are as follows:

[0058] A self-supervised learning framework is constructed based on the prediction network and the target network, and ST-GCN is used as the encoder of the prediction network and the target network.

[0059] The original human motion skeleton sequence is obtained, and the sample of each group of sequences obtained is enhanced through the behavior observation module, and then the enhanced sample data is characterized.

[0060] The original human action skeleton sequence S = {x1, x2, x3, ..., x T-2 ,x T-1 ,x T}, after random cutting and random rotation of each data enhancement and behavior observation module, a pair of enhanced sequences S is obtained q and S p , where x i ,i∈{1,2,3,…,T-2,T-1,T} represents the i-th frame skeleton in the entire sequence, T represents the number of frames in the skeleton sequence, and the prediction network and the target network use ST-GCN as encoders respectively, where the prediction network encoder fθ For prediction branch, the target network encoder f ξ For the target branch, the sequence S will be enhanced later q and S p Input the prediction network encoder f respectively θ and the target network encoder f ξ and through the formula y θ =f θ (s q ;θ) and y ξ =f ξ (s p ξ) obtains the feature vector after characterization, and then uses the momentum update method to update the target network encoder f ξ The parameter ξ of the prediction network is updated by momentum to obtain the mapping layer of the target network. The momentum update method is consistent with the target network encoder. After obtaining the feature vector, the formula z is used respectively. θ =g θ (y θ ) and z ξ =g ξ (y ξ ), the feature representation obtained is projected to a projection layer of lower dimension through two mapping layers, where the two projection layers have exactly the same structure, both consisting of a fully connected layer, a batch normalization layer, and a ReLU nonlinear activation layer, where the dimension of each layer feature is 128, and the output z of the prediction network encoder of the prediction network is θ Input prediction layer q, prediction layer q passes through o θ =q(z θ ) to obtain the output result of the prediction network, and then θ With z ξ Normalize and calculate the mean square error between the two, then optimize the mean square error between the two to bring the two features closer in the feature space to obtain the corresponding loss function L basic .

[0061] It should be further explained that the target network encoder f ξ The specific update formula of the parameter ξ is as follows:

[0062] ξ=α·ξ+(1+α)·θ

[0063] Where α represents the momentum factor, and α∈[0,1];

[0064] Loss function L basic The specific manifestations are as follows:

[0065]

[0066] Among them, Lbasic The input is obtained later by exchanging data through two network branches.

[0067] Construct a homotopy mapping function, learn the equivalence class of samples in space, and fit the continuous homotopy relationship between sample representations to build embedded features.

[0068] Specifically, suppose there are two topological spaces X and Y, and f and g are continuous mappings of X and Y respectively. If there exists a continuous mapping H:X×I→Y, such that H(x,0)=f(x),H(x,1)=g(x),then H is used as the homotopy mapping that transforms f into g,where I=[0,1].Then the human action skeleton sequence is described as a time series set of joint 3D positions. Each frame describes an action posture by coordinates. Based on the constraints implied between the coordinates, the set of all frame data is in a low-dimensional manifold in a high-dimensional space, and each sample is expressed as a discrete curve in space. The input of the prediction network and the target network is assumed to be a continuous function defined in the manifold space in time, and the continuous function is passed through the prediction network encoder f θ and the target network encoder f ξ Projected into a point in the manifold space, and then using the projection function g θ and ξ Projecting to the same space, define p∈S n For the hyperbolic embedding variable, ||p||∈[0,1), construct the homotopy function of the hyperbolic space and establish the relationship between the hyperbolic space and the projection function.

[0069] Need to be further explained, the specific calculation formula of the homology function is as follows:

[0070] Hy(x;p)=||p||g θ +(1-||p||)g ξ

[0071] In the formula, when ||p||=0, Hy(x;0)=g ξ ;and When ||p|| changes from 0 to 1, the function Hy(x; p) changes from g ξ Continuously change to g θ , this continuous change is called homotopy, expressed as Hy(x; p):g θ :g ξ .

[0072] Specifically, we select the n-dimensional Poincare model in negative curvature space as the target embedding space. For a pair of positive samples after data enhancement, we predict the network encoder f θ and the target network encoder f ξ The encoded feature vector z θ and zξ , encode the two eigenvectors and obtain h after homotopy mapping θ and h ξ As the vector to be embedded, the vector to be embedded is normalized by L2 to obtain as well as Two sets of unit vectors are then used to calculate their similarity using the inner product form and used as the norm. as well as It is embedded in the Poincare model, where λ∈(0,1), and the parameterized Poincare model embedding is defined as v=λr, located in S n In the definition Be the Mbius sum on the Poincare model, and let η,ξ∈S n Based on the defined summation formula, it can be seen that the operation does not satisfy the commutative law. Since the Poincare model is a conformal model with the property of preserving angles, based on the summation formula 1+2η,ξ>+ξ 2 and 1-η 2 From the two coefficients, we can see that if the norms of η and ξ are closer to 1, the summation result will be closer to η, and vice versa.

[0073] In addition, it should be noted that the specific form of the Poincare model is:

[0074] S n = {x∈R n |||x2<1}

[0075] The Mbius summation formula is defined as:

[0076]

[0077] The embedded features are projected into the classification probability matrix through the clustering head, and the classification probability is used as the soft label of the input data.

[0078] Specifically, the clustering head is composed of a fully connected layer, a batch normalization layer, and a ReLU nonlinear activation layer. The dimension of the internal features of each layer is 128. The constructed clustering head is used as a multi-layer perceptron. The output data decibels of the prediction network and the target network are clustered through the multi-layer perceptron to obtain the clustering matrix Y a and Y b , by using the cosine distance to measure the similarity between cluster pairs, the same cluster is considered as a positive sample pair, and different clusters are considered as negative samples, and used as soft labels for the input data, and then the loss function is used to pull the clusters apart With the exception of The distances of all clusters other than , and finally obtain the loss function at the clustering level.

[0079] It should be noted that the specific calculation formula of cosine distance is as follows:

[0080]

[0081] Among them, (i,j)∈[1,M];

[0082] The specific calculation formula of the loss function at the clustering level is as follows:

[0083]

[0084] Among them, τ C is the temperature parameter of clustering, H(Y) represents the entropy of clustering assignment probability. Adding assignment entropy can effectively prevent most clusters from being assigned to the same class.

[0085] Construct a joint optimization network and define the final loss function to train and test the self-supervised learning framework.

[0086] Specifically, after the self-supervised learning framework based on a prediction network that updates parameters through back-propagation and a target network with gradient cut-off is built, the output of the target network is used as the prediction network target, and after each round of gradient update, the new parameters are updated by exponential sliding average to update the parameters of the target network. θ and the target network encoder f ξ The two sets of feature vectors obtained are passed through the projection layer g θ (·) and g ξ (·) is transformed to the same space, and then the prediction network is connected to a multi-layer perceptron q θ (·) makes the output close to the target network, and obtains the contrastive learning loss through the MSE loss, learns the homotopy transformation through the multi-layer perceptron h(·), and uses the similarity of the two representations as the norm The features are projected onto the Poincare model, and then the Mbius sum of the Poincare model is used to represent the representative points of the sequence in the hyperbolic space, and the corresponding loss function is obtained. The loss function L Hyp Combined with the original loss function, the final loss function.

[0087] In addition, it should be further explained that the specific calculation formula of contrastive learning loss is as follows:

[0088]

[0089] Among them, z θ =g θ (f θ (x θ )),z ξ =g ξ (f ξ (xξ ));

[0090] The specific calculation formula for the representative point in the hyperbolic space is as follows:

[0091]

[0092] in, It is obtained by exchanging the input of the prediction network and the target network, and its calculation process is consistent with λ;

[0093] The specific calculation formula of the final loss function is as follows:

[0094]

[0095] L Total =L Euc +L Clu +L Hyp .

Claims

1. A cluster representation self-supervised learning method based on hyperbolic manifold homotopy embedding, characterized by: The specific steps of this learning method are as follows: S1. Build a self-supervised learning framework based on the prediction network and the target network, and use ST-GCN as the encoder of the prediction network and the target network; S2, obtaining the original human action skeleton sequence, and performing sample enhancement on each group of sequences obtained through the behavior observation module, and then performing feature characterization on the enhanced sample data; S3, construct a homotopy mapping function, learn the equivalence class of samples in space, and fit the continuous homotopy relationship between sample representations to construct embedded features; S4, project the embedded features into the classification probability matrix through the clustering head, and use the classification probability as the soft label of the input data; S5. Build a joint optimization network and define the final loss function to train and test the self-supervised learning framework.

2. The cluster representation self-supervised learning method based on hyperbolic manifold homotopy embedding according to claim 1 is characterized in that: The specific steps of characterizing the enhanced sample data described in S2 are as follows: Q1.1: Get the original human action skeleton sequence S = {x1, x2, x3, ..., x T-2 ,x T-1 ,x T }, after random cutting and random rotation of each data enhancement and behavior observation module, a pair of enhanced sequences S is obtained q and S p , where x i ,i∈{1,2,3,…,T-2,T-1,T} represents the i-th frame skeleton in the entire sequence, and T represents the number of frames in the skeleton sequence; Q1.2: The prediction network and the target network use ST-GCN as encoders respectively, where the prediction network encoder f θ For prediction branch, the target network encoder f ξ For the target branch, the sequence S will be enhanced later q and S p Input the prediction network encoder f respectively θ and the target network encoder f ξ and through the formula y θ =f θ (s q ;θ) and y ξ =f ξ (s p ξ) obtains the feature vector after characterization, and then uses the momentum update method to update the target network encoder f ξ The specific update formula of the parameter ξ is as follows: ξ=α·ξ+(1+α)·θ Where α represents the momentum factor, and α∈[0,1]; Q1.3: Update the mapping layer of the prediction network through momentum to obtain the mapping layer of the target network. The momentum update method is consistent with the target network encoder. After obtaining the feature vector, use the formula z θ =g θ (y θ ) and z ξ =g ξ (y ξ ), projecting the obtained feature representation to a projection layer of lower dimension through two mapping layers, where the two projection layers have exactly the same structure, both consisting of a fully connected layer, a batch normalization layer, and a ReLU nonlinear activation layer, where the dimension of each layer feature is 128; Q1.4: The output z of the prediction network encoder is θ Input prediction layer q, prediction layer q passes through o θ =q(z θ ) to obtain the output result of the prediction network, and then θ With z ξ Normalize and calculate the mean square error between the two, then optimize the mean square error between the two to bring the two features closer in the feature space to obtain the corresponding loss function L basic , its specific manifestations are as follows: Among them, L basic The input is obtained later by exchanging data through two network branches.

3. The cluster representation self-supervised learning method based on hyperbolic manifold homotopy embedding according to claim 2 is characterized in that: The specific steps of constructing the homotopy mapping function described in S3 are as follows: Q2.1: Assume two topological spaces X and Y, and f and g are continuous mappings of X and Y respectively. If there exists a continuous mapping H:X×I→Y such that H(x,0)=f(x), H(x,1)=g(x), then H is used as a homotopic mapping that transforms f into g, where I=[0,1]. Then the human motion skeleton sequence is described as a time series set of joint 3D positions, and each frame describes an action posture by coordinates; S2.2: Based on the constraints implied between the coordinates, the set of all frame data is placed in a low-dimensional manifold in a high-dimensional space, and each sample is represented as a discrete curve in space. The input of the prediction network and the target network is assumed to be a continuous function defined in the manifold space in time, and this continuous function is passed through the prediction network encoder f θ and the target network encoder f ξ Projected into a point in the manifold space, and then using the projection function g θ and ξ Projected into the same space; S2.3: Define p∈S n For the hyperbolic embedding variable, ||p||∈[0,1), construct the homotopy function of the hyperbolic space and establish the relationship between the hyperbolic space and the projection function. The specific calculation formula of the homotopy function is as follows: Hy(x:p)=||p||g θ +(1−||p||)g ξ In the formula, when ||p||=0, Hy(x;0)=g ξ ;and When ||p|| changes from 0 to 1, the function Hy(x; p) changes from g ξ Continuously change to g θ , this continuous change is called homotopy, expressed as Hy(x; p):g θ :g ξ .

4. The cluster representation self-supervised learning method based on hyperbolic manifold homotopy embedding according to claim 3 is characterized in that: The specific steps for constructing embedded features described in Q3 are as follows: S3.1: The n-dimensional Poincare model in negative curvature space is selected as the target embedding space, where the Poincare model is specifically expressed as: S n ={x∈R n |||x2||<1}, For a pair of positive samples after data enhancement, the prediction network encoder f θ and the target network encoder f ξ The encoded feature vector z θ and z ξ , encode the two eigenvectors and obtain h after homotopy mapping θ and h ξ As the vector to be embedded; S3.2: The vector to be embedded is normalized by L2 to obtain as well as Two sets of unit vectors are then used to calculate their similarity using the inner product form and used as the norm. as well as It is embedded in the Poincare model, where λ∈(0,1), and the parameterized Poincare model embedding is defined as v=λr, located in S n middle; S3.3: Definitions Be the Mbius sum on the Poincare model, and let η,ξ∈S n , then the summation formula is defined as: Based on the defined summation formula, it can be seen that the operation does not satisfy the commutative law. Since the Poincare model is a conformal model with the property of preserving angles, based on the summation formula 1+2<η,ξ>+||ξ|| 2 and 1-||η|| 2 From the two coefficients, we can see that if the norms of η and ξ are closer to 1, the summation result will be closer to η, and vice versa.

5. The cluster representation self-supervised learning method based on hyperbolic manifold homotopy embedding according to claim 2 is characterized in that: The specific steps described in S4 for projecting the embedded features into the classification probability matrix through the clustering head and using the classification probability as the soft label of the input data are as follows: S4.1: The clustering head is composed of a fully connected layer, a batch normalization layer, and a ReLU nonlinear activation layer. The dimension of the internal features of each layer is 128. The constructed clustering head is used as a multi-layer perceptron. The output data decibels of the prediction network and the target network are clustered through the multi-layer perceptron to obtain the clustering matrix Y a and Y b ; S4.2: Use the cosine distance to measure the similarity between cluster pairs. Consider the same cluster as a positive sample pair and different clusters as negative samples, and use them as soft labels for the input data. The specific calculation formula of the cosine distance is as follows: Among them, (i,j)∈[1,M], and then the loss function is used To pull away clusters With the exception of The distance of all clusters other than C is the temperature parameter of clustering, and finally the loss function at the clustering level is obtained as follows: Among them, H(Y) represents the entropy of clustering assignment probability. Adding assignment entropy can effectively prevent most clusters from being assigned to the same class.

6. The cluster representation self-supervised learning method based on hyperbolic manifold homotopy embedding according to claim 1, characterized in that: The specific steps of constructing the joint optimization network described in S5 and defining the final loss function to train and test the self-supervised learning framework are as follows: S5.1: After the self-supervised learning framework based on a prediction network that updates parameters through back-propagation and a target network with gradient cut-off is built, the output of the target network is used as the prediction network target, and after each round of gradient update, the new parameters are updated by exponential sliding average to update the parameters of the target network; S5.2: Pass the prediction network encoder f θ and the target network encoder f ξ The two sets of feature vectors obtained are passed through the projection layer g θ (·) and g ξ (·) is transformed to the same space, and then the prediction network is connected to a multi-layer perceptron q θ (·) Make the output close to the target network, and obtain the contrastive learning loss through MSE loss. The specific calculation formula is as follows: Among them, z θ =g θ (f θ (x θ )),z ξ =g ξ (f ξ (x ξ )); S5.3: Learn the homotopy transformation through a multilayer perceptron h(·) and use the similarity of the two representations as the norm The features are projected onto the Poincare model, and then the Mbius sum of the Poincare model is used to represent the representative points of the sequence in the hyperbolic space: in, It is obtained by exchanging the input of the prediction network and the target network. Its calculation process is consistent with λ, so the corresponding loss function is obtained as follows: L Hyp Combined with the original loss function, the final loss function is defined as: L Total =L Euc +L Clu +L Hyp .

Images