Complex valued signal processing method based on holographic converter architecture and architecture

Abstract

The invention relates to a complex valued signal processing method based on a holographic converter architecture and the architecture. Comprising the following steps: acquiring a to-be-processed complex value signal sequence, and encoding the complex value signal sequence to obtain an encoded signal sequence; holographic self-attention operation is carried out on the coded signal sequence to obtain a holographic self-attention calculation result, and the holographic self-attention operation comprises the steps that all coded signals in the coded signal sequence are projected to a query space, a key space and a value space to obtain a query signal, a key signal and a value signal; independently calculating a similarity result between the query signal and the key signal of each coded signal, and calculating to obtain a holographic self-attention calculation result according to the value signals of all coded signals, the relative phase difference between the query signal and the key signal and the similarity result; and based on a downstream task, decoding the holographic self-attention calculation result to obtain a processing result for the complex value signal sequence. By adopting the method, the analysis processing capability of the complex value signal can be improved.

Description

Technical Field

[0001] The present application relates to the field of deep learning technology for complex-valued signals, and in particular to a complex-valued signal processing method and architecture based on a holographic transformer architecture. Background Art

[0002] Existing learning models, when processing complex-valued data in radar, sonar, biomedical sensing, quantum states, and advanced communication systems, are often unable to fully explore and utilize key physical insights contained in phase information, such as path interference, coherent superposition, and fragrance synchronization, because they are designed mainly for real-valued space. This leads to poor performance of the models in complex noisy environments, phase ambiguity, and advanced signal intelligence tasks that require sophisticated physical interpretability.

[0003] Early CVNNs (Complex-Valued Neural Networks) attempted to extend neural network operations to the complex domain, using complex weights, complex activation functions, and the Wirtinger calculus for gradient backpropagation. These networks were primarily based on convolutional neural network (CNN) or recurrent neural network (RNN) topologies. However, CNN-based CVNNs are limited by their inherently local receptive field, making it difficult to capture global context and long-range phase correlations. RNN-based CVNNs, on the other hand, face sequence processing bottlenecks and gradient propagation issues, particularly when processing long sequences of complex-valued signals.

[0004] Currently, no effective solution has been proposed to address the problem of poor analytical capability of processing complex-valued signals in existing technologies. Summary of the Invention

[0005] Based on this, it is necessary to provide a complex-valued signal processing method and architecture based on a holographic converter architecture to address the above technical problems.

[0006] In a first aspect, the present application provides a complex-valued signal processing method based on a holographic converter architecture. The method comprises:

[0007] Obtaining a complex-valued signal sequence to be processed, and encoding the complex-valued signal sequence to obtain an encoded signal sequence;

[0008] Performing a holographic self-attention operation on the coded signal sequence to obtain a holographic self-attention calculation result, wherein the holographic self-attention operation includes projecting each coded signal in the coded signal sequence into query, key, and value spaces, respectively, to obtain a query signal, a key signal, and a value signal; independently calculating a similarity result between the query signal and the key signal of each coded signal; and calculating a holographic self-attention calculation result based on a relative phase difference between the value signal, the query signal, and the key signal of all coded signals, as well as the similarity result;

[0009] Based on the preset downstream tasks, the holographic self-attention calculation results are decoded to obtain the processing results for the complex-valued signal sequence.

[0010] In one embodiment, encoding the complex-valued signal sequence to obtain the encoded signal sequence includes:

[0011] Mapping the complex-valued signal sequence to the complex-valued latent semantic space to obtain a mapping result;

[0012] The preset coding matrix is ​​combined with the mapping result to obtain a coding signal sequence.

[0013] In one embodiment, independently calculating the similarity between the query signal and the key signal of each encoded signal includes:

[0014] Calculating the amplitude alignment between the query signal and the key signal, and the cosine relationship of the initial phase difference, and combining the amplitude alignment and the cosine relationship to calculate a first coupling result;

[0015] Calculating a relative phase difference between the query signal and the key signal, and calculating a second gating term based on the relative phase difference and a preset learnable parameter, wherein the value of the learnable parameter is inversely proportional to the relative phase difference;

[0016] Based on the first coupling result and the second gating term, a similarity result is obtained.

[0017] In one embodiment, a holographic self-attention calculation result is calculated based on the relative phase difference between the value signal, the query signal and the key signal of all the coded signals, and the similarity result, including:

[0018] The relative phase difference between the value signal, the query signal and the key signal is fused so that the value signal is rotated on the complex plane according to the relative phase difference to obtain a fusion result;

[0019] The holographic self-attention calculation results are obtained based on the fusion results and similarity results.

[0020] In one embodiment, the method further comprises:

[0021] Obtain a preset complex-valued signal training set; wherein the complex-valued signal training set carries a downstream task label;

[0022] The complex-valued signal training set is input into the initial holographic converter architecture for training to obtain the downstream task training results, the first loss function results and the second loss function results are calculated, and the gradients of the first loss function results and the second loss function results are back-transmitted to the initial holographic converter architecture for iterative training to generate a fully trained holographic converter architecture; wherein, the predicted training complex-valued signal sequence is reconstructed from the training holographic self-attention calculation results obtained in the training phase, and the first loss function result is calculated based on the predicted training complex-valued signal sequence and the complex-valued signal training set; the corresponding second loss function is determined based on the downstream task, and the downstream task training results and the downstream task label are substituted into the second loss function to obtain the second loss function result.

[0023] In a second aspect, the present application also provides a holographic converter architecture. The architecture includes an encoding module, a self-attention module, and a decoding module, wherein the encoding module, the self-attention module, and the decoding module are connected in sequence;

[0024] The encoding module is used to encode the input complex-valued signal sequence to obtain an encoded signal sequence;

[0025] The self-attention module is used to perform a holographic self-attention operation on the coding signal sequence to obtain a holographic self-attention calculation result, wherein the holographic self-attention operation includes projecting each coding signal in the coding signal sequence into the query, key and value spaces respectively to obtain a query signal, a key signal and a value signal, independently calculating the similarity result between the query signal and the key signal of each coding signal, and calculating the holographic self-attention calculation result based on the relative phase difference between the value signal, the query signal and the key signal of all coding signals, as well as the similarity result;

[0026] The decoding module is used to decode the output of the self-attention module based on the preset downstream task to obtain the processing result for the complex-valued signal.

[0027] In one embodiment, the holographic converter architecture further includes a decoder;

[0028] The self-attention module is used to divide the multidimensional space into multiple subspaces, and independently perform holographic self-attention operations on each subspace to obtain sub-attention calculation results, and fuse the sub-attention calculation results output by each subspace to obtain the holographic self-attention calculation results.

[0029] In one embodiment, the self-attention module further includes an attention score calculation module, a normalization module, and a coherent superposition unit; the attention score module, the normalization module, and the coherent superposition unit are connected in sequence;

[0030] The attention score calculation module is used to scale the similarity result based on a preset scaling factor to obtain a scaled result;

[0031] The normalization module is used to normalize the scaling result based on the dimension of the key signal to obtain a normalized processing result;

[0032] The coherent superposition unit is used to fuse the relative phase differences among the value signal, the query signal and the key signal to obtain a fusion result; and obtain a holographic self-attention calculation result based on the fusion result and the normalization processing result.

[0033] In one embodiment, the holographic transformer architecture further includes a complex-valued feedforward network; the complex-valued feedforward network includes a plurality of complex-valued linear transformation layers and a complex-valued activation function, and the complex-valued feedforward network is connected to the self-attention module in a one-to-one correspondence;

[0034] The complex-valued feedforward network is used to activate the output of the self-attention module.

[0035] In one embodiment, the holographic transformer architecture further includes a holographic transformation module, wherein the holographic transformation module includes multiple self-attention modules, the holographic transformation module further includes multiple layer normalization modules and multiple residual connection units, and the complex-valued feedforward network and the self-attention module are connected in a one-to-one correspondence; the layer normalization module and the self-attention module are connected to the complex-valued feedforward network, and the multiple complex-valued feedforward networks and the self-attention modules use residual connections;

[0036] The layer normalization module is used to perform layer normalization on each self-attention module or complex-valued feedforward network.

[0037] In a third aspect, the present application further provides a computer-readable storage medium having a computer program stored thereon, which, when executed by a processor, implements the following steps:

[0038] Obtaining a complex-valued signal sequence to be processed, and encoding the complex-valued signal sequence to obtain an encoded signal sequence;

[0039] Performing a holographic self-attention operation on the coded signal sequence to obtain a holographic self-attention calculation result, wherein the holographic self-attention operation includes projecting each coded signal in the coded signal sequence into query, key, and value spaces, respectively, to obtain a query signal, a key signal, and a value signal; independently calculating a similarity result between the query signal and the key signal of each coded signal; and calculating a holographic self-attention calculation result based on a relative phase difference between the value signal, the query signal, and the key signal of all coded signals, as well as the similarity result;

[0040] Based on the preset downstream tasks, the holographic self-attention calculation results are decoded to obtain the processing results for the complex-valued signal sequence.

[0041] The above-mentioned complex-valued signal processing method and architecture based on the holographic converter architecture first obtains the complex-valued signal sequence to be processed, and encodes the complex-valued signal sequence to obtain a coded signal sequence, and then performs a holographic self-attention operation on the coded signal sequence to obtain a holographic self-attention calculation result, wherein the holographic self-attention operation includes projecting each coded signal in the coded signal sequence into the query, key and value spaces respectively to obtain a query signal, a key signal and a value signal, independently calculating the similarity result between the query signal and the key signal of each coded signal, and calculating the holographic self-attention calculation result based on the relative phase difference between the value signal, the query signal and the key signal of all coded signals, and the similarity result; based on the preset downstream task, the holographic self-attention calculation result is decoded to obtain the processing result for the complex-valued signal sequence. The present application completely abandons the practice of reducing the dimensionality of the complex-valued signal or decomposing it into real-valued components for the processing method of the complex-valued signal, maintains the complex algebraic structure throughout the calculation process, and retains the amplitude and phase information integrity of the original signal to the maximum extent. The present application can improve the analytical processing capability for the complex-valued signal. BRIEF DESCRIPTION OF THE DRAWINGS

[0042] Figure 1 is an application environment diagram of a complex-valued signal processing method in one embodiment;

[0043] Figure 2 1 is a flow chart of a complex-valued signal processing method according to an embodiment;

[0044] Figure 3 A block diagram of the holographic converter architecture in one embodiment;

[0045] Figure 4 Schematic diagram of the structure of the holographic converter in a preferred embodiment. DETAILED DESCRIPTION

[0046] In order to make the purpose, technical solutions and advantages of this application more clear, the following further describes this application in detail with reference to the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are only used to explain this application and are not intended to limit this application.

[0047] The complex valued signal processing method based on the holographic converter architecture provided in the embodiment of the present application can be applied to Figure 1In the application environment shown. Among them, the terminal 102 communicates with the server 104 through the network. The data storage system can store the data that the server 104 needs to process. The data storage system can be integrated on the server 104, or it can be placed on the cloud or other network servers. First, the complex-valued signal sequence to be processed is obtained, and the complex-valued signal sequence is encoded to obtain a coded signal sequence, and a holographic self-attention operation is performed on the coded signal sequence to obtain a holographic self-attention calculation result, wherein the holographic self-attention operation includes projecting each coded signal in the coded signal sequence into the query, key and value spaces respectively to obtain a query signal, a key signal and a value signal, independently calculating the similarity result between the query signal and the key signal of each coded signal, and calculating the holographic self-attention calculation result based on the relative phase difference between the value signal, query signal and key signal of all coded signals, and the similarity result; based on the preset downstream task, the holographic self-attention calculation result is decoded to obtain a processing result for the complex-valued signal sequence. Terminal 102 may include, but is not limited to, various personal computers, laptops, smartphones, tablets, IoT devices, and portable wearable devices. IoT devices may include smart speakers, smart TVs, smart air conditioners, and smart car devices. Portable wearable devices may include smart watches, smart bracelets, and head-mounted devices. Server 104 may be implemented as a standalone server or a server cluster consisting of multiple servers.

[0048] In one embodiment, Figure 2 As shown, a complex value signal processing method based on the holographic converter architecture is provided, and the method is applied to Figure 1 The following steps are used as an example to illustrate the server in the example:

[0049] Step S210 : obtaining a complex-valued signal sequence to be processed, and performing encoding processing on the complex-valued signal sequence to obtain an encoded signal sequence.

[0050] Specifically, a complex-valued signal sequence consists of multiple complex-valued signals (tokens), each of which can be considered a complex-valued wave. This complex-valued signal sequence includes, but is not limited to, electromagnetic wave signals in radar target recognition scenarios; complex-valued analytical signals constructed through Hilbert transform (used to extract information such as instantaneous frequency or phase) in biomedical signal analysis scenarios; SAR complex images and complex signals in InSAR (Interferometric Synthetic Aperture Radar) topographic mapping scenarios; and complex-valued seismic wave signals in seismic wave analysis scenarios. Because traditional deep learning models generally focus on real-valued space, they are unable to fully explore and utilize the key physical insights contained in information such as phase. This results in poor performance in complex noise environments, phase models, and advanced signal intelligence tasks requiring sophisticated physical interpretability. The architecture in this application, however, targets complex-valued signals, understanding and utilizing their holographic nature—that is, phase and amplitude jointly encode information—to achieve efficient analysis of complex dynamic systems based on complex-valued signals. The architecture in this application simulates the coherent superposition process between the input complex-valued signal sequences, thereby learning how to decode the deep structure and information of the signal from these interference patterns.

[0051] In this embodiment, the acquired complex-valued signal sequence to be processed Perform encoding processing to obtain a coded signal sequence, where T is the sequence length, d in For the dimension of the input complex-valued features, the encoding process includes mapping the input to a higher-dimensional complex-valued latent semantic space through a complex-valued linear transformation or a complex-valued convolution operation, and performing encoding processing on it, so as to ensure that the complex-valued signal sequence to be processed maintains a clear phase information structure, and ensures that the spatiotemporal relationship and the physical nature of the complex-valued signal are encoded in a manner consistent with the physical nature of the complex-valued signal. For example, a complex-valued wavelet-based convolution encoding method can be used, and a set of preset complex-valued Morlet-Gabor wavelet filters are first used to perform multi-scale convolution on the input sequence to extract the local amplitude-phase spectrum, and then the complex-valued linear projection is performed into the complex-valued latent semantic space to obtain the above-mentioned encoded signal sequence.

[0052] Step S220, performing a holographic self-attention operation on the coding signal sequence to obtain a holographic self-attention calculation result, wherein the holographic self-attention operation includes projecting each coding signal in the coding signal sequence into the query, key and value spaces respectively to obtain a query signal, a key signal and a value signal, independently calculating the similarity results between the query signal and the key signal of each coding signal, and calculating the holographic self-attention calculation result based on the relative phase difference between the value signal, query signal and key signal of all coding signals, as well as the similarity results.

[0053] Specifically, each coded signal in the coded signal sequence is projected into the query, key and value spaces respectively to obtain the query signal, key signal and value signal, which can be transformed by three independent sets of complex-valued linear transformation matrices (parameters are W Q 、W K 、 , where these three sets of independent complex-valued linear transformation matrices are used to map the above-mentioned coded signal sequence to the three subspaces of query Q, key K, and value V. Its essence is the channel transformation matrix) Project the input complex-valued signal sequence Z to the query, key, and value spaces to obtain Q, K, , these projections aim to transform the input symbols into a specific complex-valued space in order to evaluate the potential for interference between them, and at this point we have:

[0054] Q=ZW Q 、K=ZW K 、V=ZW V ;

[0055] The above Q, K, and V are the query signal, key signal, and value signal of the above coding signal sequence as a whole, respectively. The corresponding Q i , K i and V i They represent the query signal, key signal, and value signal corresponding to a single token in the sequence, respectively. The two are just symbolic conventions of different granularities.

[0056] The similarity result between the query signal and the key signal of each coded signal in the coded signal sequence is independently calculated, wherein the similarity result is a comprehensive consideration of the amplitude and phase relationship between the query signal and the key signal of each coded signal. In some embodiments, the comprehensive similarity result between the query signal and the key signal is calculated, and the query signal Q i In the key signal K j The projection is performed in the direction, and the real part of the projection can reflect the similarity result between the query signal and the key signal.

[0057] After the similarity result is calculated, the holographic self-attention calculation result H can be obtained based on the relative phase difference between the value signal, query signal and key signal of all coded signals and the above similarity results. i , among which, specifically:

[0058]

[0059] in, For the above similarity results, V j is the value signal mentioned above, which is in vector form. The relative phase difference between the query signal and the key signal is: , Arg(·) represents the principal value of the complex argument, and T is the length of the complex signal sequence. This term indicates that before the individual coded signals in the complex-valued signal sequence are superimposed, the value vector of each coded signal will be adjusted according to the relative phase difference between the corresponding query signal and the key signal. , rotated on the complex plane, this rotation operation allows the summation process to truly simulate the constructive interference of waves (when V j The phase and adjusted phase alignment) and destructive interference (when V j The phase and This allows the network to learn complex phase-based dependencies and interference patterns in the input sequence. Through this holographic self-attention computation, the model explicitly simulates and utilizes the coherent interference phenomenon of waves within a deep learning framework, fundamentally understanding and exploiting complex signal interactions driven by phase differences.

[0060] Step S230: Based on the preset downstream task, the holographic self-attention calculation result is decoded to obtain the processing result for the complex-valued signal sequence.

[0061] Specifically, based on preset downstream tasks (such as classification tasks, regression tasks, anomaly detection, etc.), the holographic self-attention calculation results are decoded and processed to output task-related prediction results. The analysis head in the decoder of the holographic converter architecture can be designed according to the above-mentioned downstream tasks, so as to output task-related prediction results. The prediction results related to the preset downstream tasks are the above-mentioned processing results.

[0062] Through steps S210 to S230, the similarity results between the query signal and the key signal of each coded signal are calculated. The similarity results can reflect the comprehensive similarity results of the amplitude and phase relationship between the query signal and the key signal of the coded signal. The similarity results can be used as the attention weight when the coded signals are coherently superimposed in the subsequent steps, thereby amplifying attention and suppressing interference; further, based on the relative phase difference between the value signal, query signal and key signal of all coded signals, and the above similarity results, the holographic self-attention calculation results of the complex-valued signal sequence are calculated, which enables the network to learn the complex phase-based dependencies and interference patterns in the input sequence, thereby improving the deep learning model's ability to parse complex-valued signals.

[0063] In practical applications, the holographic converter architecture in this application has a wide range of application scenarios, for example, radar target recognition scenarios. In current related technologies, due to factors such as multipath and phase noise in this scenario, the target scattering center information is blurred. However, if the holographic converter architecture in this application is adopted, the ability to analyze radar signals in radar target recognition scenarios is improved, thereby improving the accuracy of target recognition. In practical applications, it is only necessary to maintain a 1-D time sequence token and convert the similarity results into The initial value of α is set to 0.3 to adapt to the rapid phase drift introduced by high-speed Doppler, so that better target recognition effect can be achieved; in the biomedical EEG (Electroencephalogram) phase-amplitude coupling detection scenario, the existing technology is difficult to measure weak phase synchronization, and if the holographic converter architecture in this application is adopted, because the analysis ability of neuronal activity potentials in medical scenarios is improved, the accuracy of classification of neuronal activity can be improved. In actual applications, only the sequence length needs to be shortened and the remaining modules can be reused to achieve better detection effect, etc. In addition to the above two application scenarios, the holographic converter architecture in this application can also be applied to phase unwrapping and terrain change monitoring in InSAR terrain mapping; and classification and prediction of collected seismic waves in seismic wave analysis.

[0064] In one embodiment, encoding the complex-valued signal sequence to obtain the encoded signal sequence includes:

[0065] Mapping the complex-valued signal sequence to the complex-valued latent semantic space to obtain a mapping result;

[0066] The preset coding matrix is ​​combined with the mapping result to obtain a coding signal sequence.

[0067] Specifically, the steps in this embodiment are preferably performed based on the encoding module of the above-mentioned holographic converter architecture. The complex-valued embedding layer in the encoding module receives the original complex-valued signal sequence X as input, and maps the above-mentioned complex-valued signal sequence X to a higher-dimensional complex-valued latent semantic space Z0∈ through complex-valued linear transformation or complex-valued convolution operation. , and the above mapping results are obtained. This process aims to completely preserve the initial amplitude and phase characteristics of each input symbol in the replica signal sequence, and lays the foundation for subsequent deep feature extraction. In some preferred embodiments, all complex-valued weight parameters are optimized using Wirtinger algorithm.

[0068] In order to enable the holographic converter architecture to perceive the order or spatial position of symbols in the sequence, a phase-consistent position coding is introduced, that is, the preset coding matrix P is integrated with the above mapping result. This method is different from the standard converter that simply adds the position coding to the real-valued embedding. By integrating the coding matrix P with the mapping result, a clear phase information structure can still be maintained, ensuring that the spatiotemporal relationship is encoded in a manner consistent with the physical nature of the complex-valued signal. For example, complex exponential coding or expansion of sine / cosine pairs can be used to ensure that the position information is effectively represented on the complex plane.

[0069] Furthermore, the encoding matrix P is generated based on the input complex-valued signal sequence, specifically: P[t,d]=r0 exp(j φ(t,d)), where φ(t,d)=2π t d / d model , and φ(t,d) is a linear phase term, which makes different time steps form equidistant phase rotation in the complex plane; r0 represents the position encoding base amplitude, r0 is greater than 0, which determines the overall amplitude calibration, d represents the feature dimension index, and 0 ≤ d < d model , so that the encoding matrix P can be obtained.

[0070] Through this embodiment, the input complex-valued signal sequence is encoded, and the initial amplitude and phase characteristics of each input symbol can be completely preserved, laying the foundation for subsequent deep feature extraction.

[0071] In one embodiment, independently calculating the similarity between the query signal and the key signal of each encoded signal includes:

[0072] Calculating the amplitude alignment between the query signal and the key signal, and the cosine relationship of the initial phase difference, and combining the amplitude alignment and the cosine relationship to calculate a first coupling result;

[0073] Calculating a relative phase difference between the query signal and the key signal, and calculating a second gating term based on the relative phase difference and a preset learnable parameter, wherein the value of the learnable parameter is inversely proportional to the relative phase difference;

[0074] Based on the first coupling result and the second gating term, a similarity result is obtained.

[0075] Specifically, this embodiment provides a specific method for calculating the similarity result between the query signal and the key signal of each coded signal:

[0076]

[0077] Among them, Q i and Kj They are the above query signal and key signal respectively, It's K j The complex conjugate of This term captures the Q i and K j The degree of amplitude alignment between them and the cosine relationship of their initial phase difference is essentially Q i In K j The real part of the projection in the direction reflects the Q i and K j The degree of "in phase" between them, and this item is the first coupling result mentioned above.

[0078] Calculate the relative phase difference between the query signal and the key signal, that is, the above , based on the relative phase difference and the preset learnable α parameter, the second gating term is calculated, that is, , where α≥0, the α parameter is a learnable scalar parameter that controls the model’s sensitivity to phase mismatch. When it is small, it means that the phase between the query signal and the key signal is highly correlated. At this time, the learnable parameter α is close to 1, allowing strong attention. When α is large, it indicates that the query signal and the key signal are phase-independent or out-of-phase. At this point, the learnable parameter α approaches 0, suppressing attention. In summary, learning the learnable parameter α enables the model to adaptively adjust its focus on the phase-correlated scale of different data features. The learnable parameter α can be set as a learnable non-negative scalar. Using the softplus(β) mapping (β is an unbounded real-valued parameter) ensures that α ≥ 0. During training, L2 regularization is applied to α to prevent it from increasing indefinitely. The initial value of α can be set to 1.0. In practical applications, α generally converges to a range of 0.1 to 3, enabling adaptive adjustment of sensitivity to phase mismatch.

[0079] In practical applications, the first coupling result and the second gating term can be directly heatmapped to observe the coherence strength between different tokens. Can be decomposed Distribution, drawing rotating arrows on the complex plane can show the constructive / destructive interference paths. Moreover, the above-mentioned knowable structures can be naturally generated in the above scheme, without requiring additional changes to the holographic transformer architecture.

[0080] This embodiment enables the model to adaptively adjust its focus on the phase correlation scale of different data features. Furthermore, the first coupling result, the second gating term, and the phase modulation term can be visualized and analyzed, revealing how the model aggregates information based on the coherence relationships between signals, providing deep, physics-based insights for understanding complex systems.

[0081] In one embodiment, a holographic self-attention calculation result is calculated based on the relative phase difference between the value signal, the query signal and the key signal of all the coded signals, and the similarity result, including:

[0082] The relative phase difference between the value signal, the query signal and the key signal is fused so that the value signal is rotated on the complex plane according to the relative phase difference to obtain a fusion result;

[0083] The holographic self-attention calculation results are obtained based on the fusion results and similarity results.

[0084] Specifically, this embodiment provides a specific holographic self-attention H i The calculation methods are as follows:

[0085] in, This is the similarity result mentioned above, V j For the above value signal, The relative phase difference between the query signal and the key signal is firstly fused to obtain the relative phase difference between the value signal, the query signal and the key signal. , which means that before the signals in the complex-valued signal sequence are superimposed and summed, the value signal V of each signal j Will be based on Rotation is performed on the complex plane. This rotation operation enables the subsequent summation process to truly simulate the constructive and destructive interference of waves, which also enables the network to learn the complex phase-based dependencies and interference patterns in the input sequence.

[0086] In one embodiment, the method further comprises:

[0087] Obtain a preset complex-valued signal training set; wherein the complex-valued signal training set carries a downstream task label;

[0088] The complex-valued signal training set is input into the initial holographic converter architecture for training to obtain the downstream task training results, the first loss function results and the second loss function results are calculated, and the gradients of the first loss function results and the second loss function results are back-transmitted to the initial holographic converter architecture for iterative training to generate a fully trained holographic converter architecture; wherein, the predicted training complex-valued signal sequence is reconstructed from the training holographic self-attention calculation results obtained in the training phase, and the first loss function result is calculated based on the predicted training complex-valued signal sequence and the complex-valued signal training set; the corresponding second loss function is determined based on the downstream task, and the downstream task training results and the downstream task label are substituted into the second loss function to obtain the second loss function result.

[0089] Specifically, this embodiment provides a method for training the above-mentioned holographic transformer architecture. In some preferred embodiments, the holographic transformer architecture is supervised training.

[0090] First, a preset complex-valued signal training set is obtained. The training set carries a downstream task label, wherein the training set is determined based on the specific downstream task. For example, if the downstream task is to predict the target distance based on the electromagnetic wave signal, then the training set includes at least the electromagnetic wave transmission signal and the echo signal. Accordingly, the above downstream task label includes at least the true target distance corresponding to the transmission signal and the echo signal in the training set. For example, in the EEG phase-amplitude coupling (PAC) detection scenario in biomedicine, if the downstream task is to perform classification prediction based on the EEG rhythm signal to determine whether a neurological disease (such as schizophrenia, Parkinson's disease, Alzheimer's disease, etc.) exists, then the training set includes at least the existing EEG rhythm signal. Accordingly, the above downstream task label includes at least the correct classification result corresponding to the EEG rhythm signal, and so on.

[0091] The complex-valued signal training set is input into the initial holographic transformer architecture for training. The downstream task training results corresponding to the training set are obtained through the analysis head in the decoder. The analysis head is designed based on the specific downstream task (such as classification, regression, anomaly detection, etc.) and outputs the prediction results corresponding to the downstream task. The first loss function result and the second loss function result are then calculated. Based on the weighted sum of the first loss function result and the second loss function result, the overall loss function result is obtained. The gradient of the overall loss function result is then transferred back to the initial holographic transformer architecture for iterative training to generate a fully trained holographic transformer architecture.

[0092] Furthermore, the following is a specific method for calculating the above-mentioned first loss function result and the second loss function result: reconstruct the predicted training complex-valued signal sequence from the training holographic self-attention calculation result calculated in the training phase (it can be understood that the calculation method of the training holographic self-attention calculation result in the training phase is the same as the calculation method of the holographic self-attention calculation result above) , the first loss function L recon It is usually the complex mean square error (complex MSE) or the joint error of amplitude and phase, for example:

[0093]

[0094] Where E is the batch expectation, Represents the real part, Represents the imaginary part, refers to the square of the norm (i.e., the sum of squares in the complex domain). This first loss function enforces that the latent representation learned by the model must retain sufficient original signal information (including fine phase structure), thereby acting as a powerful physically consistent regularizer. X is the complex-valued signal training set input to the initial holographic transformer architecture.

[0095] Then, the corresponding second loss function is determined based on the downstream task. The second loss function can be determined according to the specific downstream task type (such as cross entropy loss, root mean square error, etc.), and the prediction results related to the task are output according to the downstream task, that is, the training results of the above downstream task. , substitute the downstream task training results and the corresponding downstream task label y into the second loss function L analysis In the example above, we get the second loss function result, which is:

[0096]

[0097] Among them, TaskSpecificLoss (·) is a task-specific loss function, which is a loss function customized for specific task requirements in deep learning.

[0098] In summary, after calculating the results of the first loss function and the second loss function, the overall training goal is the weighted sum of these two losses, that is, L total =λ r L recon +λ a L analysis , where λ r and λ a is a hyperparameter used to balance the reconstruction fidelity and downstream task performance. Through the joint optimization based on the results of the first loss function and the second loss function, the model is guided to learn a complex-valued representation that is both physically meaningful and discriminative for specific tasks. Finally, based on the above joint optimization results, the loss function L is set to total The resulting gradients are back-transferred to the initial holographic transformer architecture for iterative training to generate a fully trained holographic transformer architecture.

[0099] It should be understood that, although the various steps in the flowcharts involved in the various embodiments described above are displayed in sequence according to the instructions of the arrows, these steps are not necessarily executed in sequence in the order indicated by the arrows. Unless otherwise specified herein, there is no strict order restriction on the execution of these steps, and these steps can be executed in other orders. Moreover, at least a portion of the steps in the flowcharts involved in the various embodiments described above can include multiple steps or multiple stages, and these steps or stages are not necessarily executed and completed at the same time, but can be executed at different times, and the execution order of these steps or stages is not necessarily to be carried out in sequence, but can be executed in turn or alternately with other steps or at least a portion of steps or stages in other steps.

[0100] Based on the same inventive concept, embodiments of the present application also provide a holographic converter architecture for implementing the aforementioned complex-valued signal processing method based on a holographic converter architecture. The implementation solution provided by this holographic converter architecture is similar to the implementation solution described in the aforementioned method. Therefore, the specific limitations in one or more holographic converter architecture embodiments provided below can be found in the above-mentioned limitations on the complex-valued signal processing method based on a holographic converter architecture, and will not be repeated here.

[0101] In one embodiment, Figure 3 As shown, a holographic converter architecture is provided, including: an encoding module 31, a self-attention module 32 and a decoding module 33, wherein the encoding module 31, the self-attention module 32 and the decoding module 33 are connected in sequence:

[0102] The encoding module 31 is used to encode the input complex-valued signal sequence to obtain a coded signal sequence;

[0103] The self-attention module 32 is used to perform a holographic self-attention operation on the coded signal sequence to obtain a holographic self-attention calculation result, wherein the holographic self-attention operation includes projecting each coded signal in the coded signal sequence into the query, key and value spaces respectively to obtain a query signal, a key signal and a value signal, independently calculating the similarity result between the query signal and the key signal of each coded signal, and calculating the holographic self-attention calculation result based on the relative phase difference between the value signal, the query signal and the key signal of all the coded signals and the similarity result;

[0104] The decoding module 33 is used to decode the output of the self-attention module based on a preset downstream task to obtain a processing result for the complex-valued signal.

[0105] In one embodiment, the holographic converter architecture further includes a decoder;

[0106] The self-attention module is used to divide the multidimensional space into multiple subspaces, and independently perform holographic self-attention operations on each subspace to obtain sub-attention calculation results, and fuse the sub-attention calculation results output by each subspace to obtain the holographic self-attention calculation results.

[0107] Specifically, the self-attention module divides the multidimensional space into multiple subspaces (also called "multi-heads") and performs the above-mentioned holographic self-attention operation independently on each subspace, obtaining the above-mentioned sub-attention calculation results on each subspace. This allows the model to pay attention to information from different representation subspaces with different interference patterns in parallel. The outputs of each head are then spliced, and the sub-attention calculation results output by each subspace are fused to obtain the above-mentioned holographic self-attention calculation results. Among them, the dimension of the multidimensional space is d model The dimension of the multidimensional space is equal to the dimension of the query signal matrix, key signal matrix and value signal matrix of the complex-valued signal sequence mentioned above.

[0108] In summary, it can be understood that this embodiment proposes another method for calculating the holographic self-attention calculation results, namely, independently performing the holographic self-attention operation described above on multiple subspaces, and fusing the sub-attention calculation results output by each subspace to obtain the final holographic self-attention calculation result. Compared with performing the holographic self-attention operation in one space in the above text, it can more effectively capture the diversity of data and enhance the model's recognition ability of complex data.

[0109] In one embodiment, the self-attention module further includes an attention score calculation module, a normalization module, and a coherent superposition unit; the attention score module, the normalization module, and the coherent superposition unit are connected in sequence;

[0110] The attention score calculation module is used to scale the similarity result based on a preset scaling factor to obtain a scaled result;

[0111] The normalization module is used to normalize the scaling result based on the dimension of the key signal to obtain a normalized processing result;

[0112] The coherent superposition unit is used to fuse the relative phase differences among the value signal, the query signal and the key signal to obtain a fusion result; and obtain a holographic self-attention calculation result based on the fusion result and the normalization processing result.

[0113] Specifically, after the similarity result is calculated, the similarity result is input into the attention score calculation module. The attention score calculation module scales the similarity result by a preset scaling factor to obtain the scaled result. At this time, there is:

[0114] where dk This is the scaling factor mentioned above, which is similar to the standard transformer and is used to stabilize the gradient.

[0115] Furthermore, after obtaining the scaling result, the scaling result is input into the normalization module. The normalization module normalizes the scaling result based on the dimension of the key signal to obtain the normalized processing result. At this time, there is:

[0116]

[0117] A ij This is the normalization result.

[0118] Finally, based on the relative phase difference between the value signal, the query signal, and the key signal, a fusion process is performed to obtain the fusion result, and based on the fusion result and the normalization processing result, the above holographic self-attention calculation result is obtained. At this time, we have:

[0119] Compared with the above, the similarity result is combined with the fusion result. In this embodiment, the holographic self-attention calculation result is obtained based on the fusion result and the normalization processing result, which not only ensures the stability of the gradient, Prevent d k When the gradient is very large, it explodes, and normalization is performed before combining the two. Normalization can avoid a few maximum values ​​dominating the training, which can effectively improve the convergence speed, numerical stability and performance of the model.

[0120] In one embodiment, the holographic transformer architecture further includes a complex-valued feedforward network; the complex-valued feedforward network includes a plurality of complex-valued linear transformation layers and a complex-valued activation function, and the complex-valued feedforward network is connected to the self-attention module in a one-to-one correspondence;

[0121] The complex-valued feedforward network is used to activate the output of the self-attention module.

[0122] Specifically, the holographic converter architecture includes a complex feed-forward network (CFFN), which is preferably composed of two complex-valued linear transformation layers and a complex-valued activation function therebetween. The complex-valued activation function includes but is not limited to a concatenated rectified linear unit (CReLU), a complex rectified linear unit (zReLU), and a modified rectified linear unit (modReLU).

[0123] The complex-valued feedforward network and the self-attention module are connected one-to-one. The complex-valued feedforward network has:

[0124] CFFN(x)=ComplexLinear2(CActivation(ComplexLinear1(x)));

[0125] ComplexLinear1(·) and ComplexLinear2(·) both perform linear transformations on the input parameters in the complex domain, equivalent to fully connected layers in the real domain, but their weights and biases are all complex. CActivation(·) applies a nonlinear activation to the output of the complex linear layer of the input parameters. In summary, the two-layer ComplexLinear + CActivation can be regarded as a complex-valued feedforward network, whose weights and biases are all complex.

[0126] In summary, the above complex-valued feedforward network can realize signal feature extraction and information enhancement.

[0127] In one embodiment, the architecture further includes a holographic transformation module, wherein the holographic transformation module includes multiple self-attention modules, the holographic transformation module further includes multiple layer normalization modules and multiple residual connection units, and the complex-valued feedforward network and the self-attention module are connected one-to-one; the layer normalization module and the self-attention module are connected to the complex-valued feedforward network, and the multiple complex-valued feedforward networks and the self-attention module use residual connections;

[0128] The layer normalization module is used to perform layer normalization on each self-attention module or complex-valued feedforward network.

[0129] Specifically, the holographic transformer module includes multiple self-attention modules, and the self-attention modules and the complex-valued feedforward networks are connected one-to-one. The layer normalization module is set before the input of each self-attention module, or before the input of each complex-valued feedforward network, so that the complex-valued version of layer normalization is used to stabilize the training process. Residual connections are used around each sub-layer (self-attention module or complex-valued feedforward network). By stacking multiple such holographic transformer modules, a deep model can be constructed to learn complex hierarchical complex-valued features in the data.

[0130] Each module in the holographic converter architecture described above can be implemented in whole or in part through software, hardware, or a combination thereof. Each module can be embedded in or independent of a processor in a computer device in hardware form, or can be stored in a computer device memory in software form, so that the processor can call and execute the corresponding operations of each module.

[0131] This application also provides a preferred embodiment of a holographic converter architecture. Figure 4Schematic diagram of the structure of the holographic converter in a preferred embodiment.

[0132] First, a complex-valued signal sequence to be processed is obtained. Based on the complex-valued embedding layer in the encoding module, a complex-valued linear transformation or a complex-valued convolution operation is performed on the complex-valued signal sequence to map the complex-valued signal sequence to a higher-dimensional complex-valued latent semantic space to obtain a mapping result. The mapping result is then combined with a preset encoding matrix to obtain a coded signal sequence. This allows the coded signal sequence to maintain a clear phase information structure and ensures that the spatiotemporal relationship is encoded in a manner consistent with the physical nature of the complex-valued signal. The coded signal sequence Z can be expressed as: Z = Z0 + P, where Z0 is the mapping result and P is the preset encoding matrix, which is in complex exponential form.

[0133] After obtaining the coded signal sequence, the coded signal sequence is input into the above-mentioned holographic transformation module. The holographic transformation module includes the above-mentioned multi-head self-attention module, complex-valued feedforward network, layer normalization module and residual connection unit, wherein the multi-head self-attention module divides the multi-dimensional space into multiple subspaces (i.e., multiple "heads"), and independently performs holographic self-attention operations on each subspace. Each subspace outputs a corresponding sub-attention calculation result, and then the self-attention calculation results output by each subspace are spliced ​​to obtain the above-mentioned holographic self-attention calculation result, and the self-attention result is input into the complex-valued feedforward network (CFFN). The holographic self-attention calculation result is activated by the complex-valued feedforward network, and a complex-valued version of the layer normalization module can be connected before the input end of each complex-valued feedforward network or self-attention module to stabilize the training process, and residual connections can be used around multiple complex-valued feedforward networks or self-attention modules to build a deep model to learn complex hierarchical replication features in the data. It should be further explained that the above holographic self-attention operation includes first projecting each coded signal in the coded signal sequence into the query, key and value spaces respectively to obtain the query signal, key signal and value signal, and then independently calculating the similarity between the query signal and the key signal of each coded signal. , and then the similarity result is scaled by the attention score calculation module to obtain the scaled result , normalize the scaling result based on the key signal dimension to obtain the normalized result Finally, the relative phase differences between the value signal, query signal and key signal of all the coded signals in the coded signal sequence are fused to obtain the fusion result, and the above holographic self-attention calculation result H is obtained based on the fusion result and the normalization result. i , where specifically, calculate H i have:

[0134] Its core lies in This term means that before the coded signals in the coded signal sequence are superimposed, the value vector V of each coded signal j According to the corresponding query signal Q i The relative phase difference between Rotation is performed on the complex plane. This rotation operation enables the summation process to realistically simulate constructive and destructive interference between waves, allowing the network to learn complex phase-based dependencies and interference patterns in the input sequence.

[0135] The holographic self-attention calculation results are input into the decoding module, which includes the reconstruction head and the analysis head. During the training phase, the first loss function result is calculated by the reconstruction head. , calculate the second loss function result L by analyzing the head analysis , and calculate the weighted sum of the first loss function result and the second loss function result through joint training, and transmit the gradient of the weighted sum result back to the initial holographic transformer architecture for iterative training to generate a fully trained holographic transformer architecture; and in the use stage after the training is complete, only the analysis head can be used to output the prediction result corresponding to the downstream task through the analysis head, that is, the processing result described above In other embodiments, a reconstruction head can also be used during the usage phase to reconstruct a complex-valued signal sequence from the holographic self-attention calculation results, and the reconstructed complex-valued signal sequence can be compared with the original input complex-valued signal sequence to detect the performance of the holographic converter architecture.

[0136] In some preferred embodiments, phase smoothing constraints, amplitude-phase decoupling regularization, dedicated complex-valued initialization, and Wirtinger optimizer adaptation may also be introduced. The phase smoothing constraint is to promote the learning of a locally smooth phase representation of a complex-valued signal sequence in the time domain or spatial domain. A regularization term can be introduced to penalize excessive phase gradients between adjacent complex-valued signals (tokens) in the complex-valued signal sequence, which helps to reduce the impact of entanglement and noise on model performance. The amplitude-phase decoupling regularization is to introduce auxiliary losses to encourage the information encoded by the amplitude and phase channels in the intermediate layer representation to be decoupled or orthogonal to a certain extent, thereby preventing the model from over-relying on a single information channel or producing trivial solutions. The dedicated complex-valued initialization (Specialized Complex Weight Initialization) uses initialization methods such as real and imaginary parts based on independent Gaussian distributions, or initialization methods that keep the weight matrix satisfying specific complex norm constraints, which is crucial for stable training of deep replica networks. The Wirtinger optimizer adaptation is that the optimizer (such as Adam) needs to be correctly configured to process the complex-valued gradients calculated based on the Wirtinger calculus.

[0137] Based on the same inventive concept, the present application also provides a computer-readable storage medium having a computer program stored thereon. When the computer program is executed by a processor, the following steps are implemented:

[0138] Obtaining a complex-valued signal sequence to be processed, and encoding the complex-valued signal sequence to obtain an encoded signal sequence;

[0139] Performing a holographic self-attention operation on the coded signal sequence to obtain a holographic self-attention calculation result, wherein the holographic self-attention operation includes projecting each coded signal in the coded signal sequence into query, key, and value spaces, respectively, to obtain a query signal, a key signal, and a value signal; independently calculating a similarity result between the query signal and the key signal of each coded signal; and calculating a holographic self-attention calculation result based on a relative phase difference between the value signal, the query signal, and the key signal of all coded signals, as well as the similarity result;

[0140] Based on the preset downstream tasks, the holographic self-attention calculation results are decoded to obtain the processing results for the complex-valued signal sequence.

[0141] It should be noted that the user information (including but not limited to user device information, user personal information, etc.) and data (including but not limited to data used for analysis, stored data, displayed data, etc.) involved in this application are all information and data authorized by the user or fully authorized by all parties.

[0142] Those skilled in the art will appreciate that all or part of the processes in the above-mentioned embodiments can be implemented by instructing the relevant hardware through a computer program. The computer program can be stored in a non-volatile computer-readable storage medium. When the computer program is executed, it can include the processes of the above-mentioned embodiments. In particular, any reference to memory, database, or other media used in the embodiments provided in this application can include at least one of non-volatile and volatile memory. Non-volatile memory can include read-only memory (ROM), magnetic tape, floppy disk, flash memory, optical memory, high-density embedded non-volatile memory, resistive random access memory (ReRAM), magnetic random access memory (MRAM), ferroelectric random access memory (FRAM), phase change memory (PCM), graphene memory, etc. Volatile memory can include random access memory (RAM) or external cache memory, etc. By way of illustration and not limitation, RAM can take various forms, such as static random access memory (SRAM) or dynamic random access memory (DRAM). The databases involved in the various embodiments provided herein may include at least one of a relational database and a non-relational database. Non-relational databases may include, but are not limited to, distributed databases based on blockchains. The processors involved in the various embodiments provided herein may be, but are not limited to, general-purpose processors, central processing units (CPUs), graphics processing units (GPUs), digital signal processors (DSPs), programmable logic devices (PLDs), data processing logic devices based on quantum computing, and the like.

[0143] The technical features of the above embodiments can be combined arbitrarily. To make the description concise, not all possible combinations of the technical features in the above embodiments are described. However, as long as there is no contradiction in the combination of these technical features, they should be considered to be within the scope of this specification.

[0144] The above-described embodiments merely represent several implementation methods of the present application. While the descriptions are relatively specific and detailed, they should not be construed as limiting the scope of the present application. It should be noted that a person of ordinary skill in the art may make various modifications and improvements without departing from the spirit of the present application, and these modifications and improvements fall within the scope of protection of the present application. Therefore, the scope of protection of the present application shall be determined by the appended claims.

Claims

1. A complex-valued signal processing method based on a holographic converter architecture, characterized in that: The method comprises: Acquiring a complex-valued signal sequence to be processed, and performing encoding processing on the complex-valued signal sequence to obtain an encoded signal sequence; performing a holographic self-attention operation on the coded signal sequence to obtain a holographic self-attention calculation result, wherein the holographic self-attention operation includes projecting each coded signal in the coded signal sequence into query, key, and value spaces, respectively, to obtain a query signal, a key signal, and a value signal; independently calculating a similarity result between the query signal and the key signal of each coded signal; and calculating the holographic self-attention calculation result based on a relative phase difference between the value signal, the query signal, and the key signal of all the coded signals, and the similarity result; Based on the preset downstream tasks, the holographic self-attention calculation results are decoded to obtain the processing results for the complex-valued signal sequence.

2. The method according to claim 1, characterized in that The encoding process is performed on the complex-valued signal sequence to obtain an encoded signal sequence, comprising: Mapping the complex-valued signal sequence to a complex-valued latent semantic space to obtain a mapping result; The preset coding matrix is ​​combined with the mapping result to obtain the coding signal sequence.

3. The method according to claim 1, characterized in that The independently calculating the similarity results between the query signal and the key signal of each of the coded signals includes: Calculating an amplitude alignment degree between the query signal and the key signal, and a cosine relationship of an initial phase difference, and calculating a first coupling result by combining the amplitude alignment degree and the cosine relationship; Calculating a relative phase difference between the query signal and the key signal, and obtaining a second gating term based on the relative phase difference and a preset learnable parameter, wherein a value of the learnable parameter is inversely proportional to the relative phase difference; The similarity result is obtained based on the first coupling result and the second gating item.

4. The method according to claim 1, wherein The holographic self-attention calculation result is calculated based on the value signal of all the coded signals, the relative phase difference between the query signal and the key signal, and the similarity result, including: fusing the relative phase difference between the value signal, the query signal, and the key signal so that the value signal is rotated on a complex plane according to the relative phase difference to obtain a fusion result; The holographic self-attention calculation result is obtained based on the fusion result and the similarity result.

5. The method according to claim 1, wherein The method further comprises: Obtaining a preset complex-valued signal training set; wherein the complex-valued signal training set carries a downstream task label; The complex-valued signal training set is input into the initial holographic converter architecture for training to obtain the downstream task training result, the first loss function result and the second loss function result are calculated, and the gradients of the first loss function result and the second loss function result are reversely transmitted to the initial holographic converter architecture for iterative training to generate the fully trained holographic converter architecture; wherein, the predicted training complex-valued signal sequence is reconstructed from the training holographic self-attention calculation result obtained in the training stage, and the first loss function result is calculated based on the predicted training complex-valued signal sequence and the complex-valued signal training set; the corresponding second loss function is determined based on the downstream task, and the downstream task training result and the downstream task label are substituted into the second loss function to obtain the second loss function result.

6. A holographic converter architecture, characterized in that: The holographic converter architecture includes an encoding module, a self-attention module and a decoding module, wherein the encoding module, the self-attention module and the decoding module are connected in sequence; The encoding module is used to encode the input complex-valued signal sequence to obtain an encoded signal sequence; The self-attention module is used to perform a holographic self-attention operation on the coded signal sequence to obtain a holographic self-attention calculation result, wherein the holographic self-attention operation includes projecting each coded signal in the coded signal sequence into query, key and value spaces respectively to obtain a query signal, a key signal and a value signal, independently calculating a similarity result between the query signal and the key signal of each coded signal, and calculating the holographic self-attention calculation result based on the relative phase difference between the value signal, the query signal and the key signal of all the coded signals, and the similarity result; The decoding module is used to decode the output of the self-attention module based on a preset downstream task to obtain a processing result for the complex-valued signal.

7. The architecture according to claim 6, characterized in that The holographic converter architecture also includes a decoder; The self-attention module is used to divide the multidimensional space into multiple subspaces, and independently perform the holographic self-attention operation on each subspace to obtain the sub-attention calculation results, and fuse the sub-attention calculation results output by each subspace to obtain the holographic self-attention calculation results.

8. The architecture according to claim 6, wherein: The self-attention module further includes an attention score calculation module, a normalization module and a coherent superposition unit; the attention score calculation module, the normalization module and the coherent superposition unit are connected in sequence; The attention score calculation module is used to scale the similarity result based on a preset scaling factor to obtain a scaling result; The normalization module is used to perform normalization processing on the scaling result based on the dimension of the key signal to obtain a normalized processing result; The coherent superposition unit is used to fuse the relative phase differences between the value signal, the query signal and the key signal to obtain a fusion result; And based on the fusion result and the normalization processing result, the holographic self-attention calculation result is obtained.

9. The architecture according to claim 6, wherein: The holographic converter architecture further includes a complex-valued feedforward network; the complex-valued feedforward network includes a plurality of complex-valued linear transformation layers and a complex-valued activation function, and the complex-valued feedforward network is connected to the self-attention module in a one-to-one correspondence; The complex-valued feedforward network is used to activate the output of the self-attention module.

10. The architecture according to claim 9, characterized in that The architecture further includes a holographic transformation module, wherein the holographic transformation module includes multiple self-attention modules, the holographic transformation module also includes multiple layer normalization modules and multiple residual connection units, and the complex-valued feedforward network and the self-attention module are connected in a one-to-one correspondence; the layer normalization module and the self-attention module are connected to the complex-valued feedforward network, and residual connections are used between the multiple complex-valued feedforward networks and the self-attention modules; The layer normalization module is used to perform layer normalization processing on each of the self-attention modules or the complex-valued feedforward network.

11. A computer-readable storage medium having a computer program stored thereon, characterized in that: When the computer program is executed by a processor, the steps of the method according to any one of claims 1 to 5 are implemented.

Images