A complex-valued signal processing method and architecture based on a holographic converter architecture

Abstract

This application relates to a complex-valued signal processing method and architecture based on a holographic transformer architecture. It includes: acquiring a sequence of complex-valued signals to be processed; encoding the sequence to obtain an encoded signal sequence; performing a holographic self-attention operation on the encoded signal sequence to obtain a holographic self-attention calculation result. The holographic self-attention operation includes projecting each encoded signal in the encoded signal sequence onto a query, key, and value space respectively to obtain a query signal, a key signal, and a value signal; independently calculating the similarity results between the query and key signals of each encoded signal; calculating the holographic self-attention calculation result based on the relative phase difference between the value signals, query signals, and key signals of all encoded signals, and the similarity results; and decoding the holographic self-attention calculation result based on a downstream task to obtain the processing result for the complex-valued signal sequence. This method can improve the analytical processing capability for complex-valued signals.

Description

Technical Field

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

[0002] Existing learning models, when dealing with complex-valued data in systems such as radar, sonar, biomedical sensing, quantum states, and advanced communication, often fail to fully explore and utilize the key physical insights contained in the phase information, such as path interference, coherent superposition, and scent synchronization, because their design is mainly aimed at the real-valued space. This results in poor performance of the models in complex noisy environments, phase ambiguity, and advanced signal intelligence tasks that require fine physical interpretability.

[0003] Early CVNNs (Complex-Valued Neural Networks) attempted to extend the computation of neural networks to the complex domain, using complex weights, complex activation functions, and Wirtinger calculus for gradient backpropagation. These networks were primarily based on the topologies of Convolutional Neural Networks (CNNs) or Recurrent Neural Networks (RNNs). However, CNN-based CVNNs are limited by their inherent local receptive field, making it difficult to capture global contextual information and long-range phase correlations. RNN-based CVNNs, on the other hand, face bottlenecks in sequence processing and gradient propagation, especially when dealing with long sequences of complex-valued signals.

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

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

[0006] Firstly, this application provides a complex-valued signal processing method based on a holographic converter architecture. The method includes:

[0007] Obtain the complex-valued signal sequence to be processed, and encode the complex-valued signal sequence to obtain the encoded signal sequence;

[0008] A holographic self-attention operation is performed on the encoded signal sequence to obtain the holographic self-attention calculation result. The holographic self-attention operation includes projecting each encoded signal in the encoded signal sequence onto the query, key, and value spaces respectively to obtain the query signal, key signal, and value signal; independently calculating the similarity results between the query signal and key signal of each encoded 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 encoded signals, as well as the similarity results.

[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, the complex-valued signal sequence is encoded to obtain an encoded signal sequence, including:

[0011] The complex-valued signal sequence is mapped to the complex-valued latent semantic space to obtain the mapping result;

[0012] The preset encoding matrix is ​​fused with the mapping result to obtain the encoded signal sequence.

[0013] In one embodiment, the similarity results between the query signal and the key signal of each encoded signal are calculated independently, including:

[0014] Calculate the amplitude alignment between the query signal and the key signal, and the cosine relationship with the initial phase difference. Combine the amplitude alignment and the cosine relationship to calculate the first coupling result.

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

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

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

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

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

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

[0021] Obtain a pre-defined complex-valued signal training set; wherein the complex-valued signal training set carries downstream task labels;

[0022] The complex-valued signal training set is input into the initial holographic transformer architecture for training to obtain the downstream task training results. 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 backpropagated to the initial holographic transformer architecture for iterative training to generate a fully trained holographic transformer architecture. Specifically, 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 downstream task labels are substituted into the second loss function to obtain the second loss function result.

[0023] Secondly, this application also provides a holographic converter architecture. The architecture includes an encoding module, a self-attention module, and a decoding module, which are connected sequentially.

[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 holographic self-attention operations on the encoded signal sequence to obtain holographic self-attention calculation results. The holographic self-attention operation includes projecting each encoded signal in the encoded signal sequence onto the query, key, and value spaces to obtain query signals, key signals, and value signals, independently calculating the similarity results between the query signals and key signals of each encoded signal, and calculating the holographic self-attention calculation results based on the relative phase difference between the value signals, query signals, and key signals of all encoded signals, as well as the similarity results.

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

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

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

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

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

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

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

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

[0034] Complex-valued feedforward networks are used to activate the output of the self-attention module.

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

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

[0037] Thirdly, this application also provides a computer-readable storage medium having a computer program stored thereon, wherein the computer program, when executed by a processor, performs the following steps:

[0038] Obtain the complex-valued signal sequence to be processed, and encode the complex-valued signal sequence to obtain the encoded signal sequence;

[0039] A holographic self-attention operation is performed on the encoded signal sequence to obtain the holographic self-attention calculation result. The holographic self-attention operation includes projecting each encoded signal in the encoded signal sequence onto the query, key, and value spaces respectively to obtain the query signal, key signal, and value signal; independently calculating the similarity results between the query signal and key signal of each encoded 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 encoded signals, as well as the similarity results.

[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 aforementioned complex-valued signal processing method and architecture based on a holographic converter architecture first acquires the complex-valued signal sequence to be processed and encodes it to obtain an encoded signal sequence. Then, a holographic self-attention operation is performed on the encoded signal sequence to obtain the holographic self-attention calculation result. The holographic self-attention operation includes projecting each encoded signal in the encoded signal sequence onto a query, key, and value space respectively to obtain a query signal, a key signal, and a value signal; independently calculating the similarity results between the query and key signals of each encoded signal; and calculating the holographic self-attention calculation result based on the relative phase difference between the value signals, query signals, and key signals of all encoded signals, as well as the similarity results. Based on a preset downstream task, the holographic self-attention calculation result is decoded to obtain the processing result for the complex-valued signal sequence. This application's method for processing complex-valued signals completely abandons the practice of reducing the dimensionality of complex-valued signals or decomposing them into real-valued components. It maintains the complex algebraic structure throughout the entire calculation process, maximizing the preservation of the amplitude and phase information integrity of the original signal. This application improves the analytical processing capability for complex-valued signals. Attached Figure Description

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

[0043] Figure 2 This is a flowchart illustrating a complex-valued signal processing method in one embodiment;

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

[0045] Figure 4 This is a schematic diagram of the holographic converter architecture in a preferred embodiment. Detailed Implementation

[0046] To make the objectives, technical solutions, and advantages of this application clearer, the following detailed description is provided in conjunction with the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are merely illustrative and not intended to limit the scope of this application.

[0047] The complex-valued signal processing method based on a holographic converter architecture provided in this application can be applied to, for example... Figure 1In the application environment shown, terminal 102 communicates with server 104 via a network. A data storage system can store the data that server 104 needs to process. The data storage system can be integrated onto server 104 or placed on a cloud or other network server. First, the complex-valued signal sequence to be processed is acquired and encoded to obtain an encoded signal sequence. A holographic self-attention operation is then performed on the encoded signal sequence to obtain the holographic self-attention calculation result. The holographic self-attention operation includes projecting each encoded signal in the encoded signal sequence onto a query, key, and value space to obtain a query signal, a key signal, and a value signal. The similarity results between the query signal and key signal of each encoded signal are calculated independently. Based on the relative phase difference between the value signal, query signal, and key signal of all encoded signals, and the similarity results, the holographic self-attention calculation result is calculated. Based on a preset downstream task, the holographic self-attention calculation result is decoded to obtain the processing result for the complex-valued signal sequence. The terminal 102 can be, but is not limited to, various personal computers, laptops, smartphones, tablets, IoT devices, and portable wearable devices. IoT devices can include smart speakers, smart TVs, smart air conditioners, and smart in-vehicle systems. Portable wearable devices can include smartwatches, smart bracelets, and head-mounted devices. The server 104 can be implemented using a standalone server or a server cluster consisting of multiple servers.

[0048] In one embodiment, such as Figure 2 As shown, a complex-valued signal processing method based on a holographic converter architecture is provided, which is then applied to... Figure 1 Taking the server in the example, the following steps are included:

[0049] Step S210: Obtain the complex-valued signal sequence to be processed, and encode the complex-valued signal sequence to obtain the encoded signal sequence.

[0050] Specifically, the 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 identification scenarios; complex-valued analytical signals constructed using Hilbert transform in biomedical signal analysis scenarios (used to extract instantaneous frequency or phase information); 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, etc. Because traditional deep learning models generally focus on real-valued spaces, they cannot fully mine and utilize the key physical insights contained in phase and other information. This leads to poor performance in complex noisy environments, phase models, and advanced signal intelligence tasks requiring fine physical interpretability. The architecture in this application, however, targets complex-valued signals, understanding and utilizing their holographic nature—that is, the co-encoding of phase and amplitude 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 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 Encoding processing is performed to obtain an encoded signal sequence, where T is the sequence length and d is the coded signal length. in To determine 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 complex-valued linear transformation or complex-valued convolution operations, and then encoding it to ensure that the spatiotemporal relationships and the physical nature of the complex-valued signal are encoded while maintaining a clear phase information structure. For example, a complex-valued wavelet basis convolutional encoding method can be used. First, a set of preset complex-valued Morlet-Gabor wavelet filters is used to perform multi-scale convolution on the input sequence to extract the local amplitude-phase spectrum, and then the sequence is entered into the complex-valued latent semantic space through complex-valued linear projection to obtain the above-mentioned encoded signal sequence.

[0052] Step S220: Perform a holographic self-attention operation on the encoded signal sequence to obtain the holographic self-attention calculation result. The holographic self-attention operation includes projecting each encoded signal in the encoded signal sequence onto the query, key, and value spaces respectively to obtain the query signal, key signal, and value signal; independently calculating the similarity results between the query signal and key signal of each encoded 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 encoded signals, as well as the similarity results.

[0053] Specifically, each encoded signal in the encoded signal sequence is projected onto the query, key, and value spaces respectively to obtain the query signal, key signal, and value signal. This can be achieved through three independent complex-valued linear transformation matrices (with parameters W...). Q W K , Among them, these three independent complex-valued linear transformation matrices are used to map the above-mentioned coded signal sequence to three subspaces: query Q, key K, and value V. Essentially, these are channel transformation matrices. The input complex-valued signal sequence Z is projected onto the query, key, and value spaces to obtain Q, K, and V. These projections are designed to transform the input symbols to a specific complex-valued space in order to assess their interference potential, and at this point:

[0054] Q=ZW Q K=ZW K V=ZW V ;

[0055] Q, K, and V above represent the query signal, key signal, and value signal of the entire encoded signal sequence, respectively, and the corresponding Q... i K i and V i These represent the query signal, key signal, and value signal corresponding to a single token in the sequence, respectively; the two are merely symbolic conventions with different granularities.

[0056] The similarity results between the query signal and the key signal of each coded signal in the coded signal sequence are calculated independently. 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 can be used as a reference. i In the key signal K j The projection is performed in the direction of the query signal and the key signal. The real part of the projection reflects the similarity result between the query signal and the key signal.

[0057] After calculating the similarity results, the holographic self-attention calculation result H can be obtained by combining the relative phase differences between the value signals, query signals, and key signals of all encoded signals with the above similarity results. i Specifically, these include:

[0058]

[0059] in, For the above similarity results, V j Given the above value signal, which is in vector form, the relative phase difference between the query signal and the key signal is: Arg(·) represents taking the principal argument of a complex number, and T is the sequence length of the complex-valued signal sequence. This indicates that, before the individual encoded signals in the complex-valued signal sequence are superimposed, the value vector of each encoded signal is determined based on the relative phase difference between its corresponding query signal and key signal. Rotation is performed on the complex plane, and this rotation operation allows the summation process to realistically simulate the constructive interference of waves (when V...). j phase and (when the phase is aligned after adjustment) and destructive interference (when V) j phase and When the adjusted phase is anti-aligned, the network can learn complex phase-based dependencies and interferograms in the input sequence. Through the above holographic self-attention computation, the model can explicitly simulate and utilize the coherent interference of waves within the deep learning framework, thereby fundamentally understanding and utilizing complex signal interactions driven by phase differences.

[0060] Step S230: Based on the preset downstream task, decode the holographic self-attention calculation result to obtain the processing result for the complex value 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 the task-related prediction results are output. The analysis head in the decoder of the holographic converter architecture can be designed according to the aforementioned downstream tasks, thereby outputting task-related prediction results. The prediction results related to the preset downstream tasks are the processing results mentioned above.

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

[0063] In practical applications, the holographic converter architecture of this application has a wide range of applications, such as radar target recognition. Current related technologies often suffer from blurred target scattering center information due to multipath propagation and phase noise in such scenarios. However, by adopting the holographic converter architecture of this application, the accuracy of target recognition is improved because it enhances the ability to analyze radar signals in radar target recognition scenarios. In practical applications, only the 1-D time-series token needs to be maintained, and the similarity results... The initial value of α is set to 0.3 to accommodate the rapid phase drift introduced by high-speed Doppler, thus achieving better target recognition results. In biomedical EEG (Electroencephalogram) phase-amplitude coupling detection scenarios, existing technologies struggle to measure weak phase synchronization. However, the holographic converter architecture described in this application improves the ability to analyze neuronal activity potentials in medical scenarios, thereby enhancing the accuracy of neuronal activity classification. In practical applications, only the sequence length needs to be shortened, and the remaining modules can be reused to achieve better detection results. In addition to the two application scenarios mentioned above, the holographic converter architecture described in this application can also be applied to phase unwrapping and terrain change monitoring in InSAR topographic mapping, as well as classification and prediction of acquired seismic waves in seismic wave analysis.

[0064] In one embodiment, the complex-valued signal sequence is encoded to obtain an encoded signal sequence, including:

[0065] The complex-valued signal sequence is mapped to the complex-valued latent semantic space to obtain the mapping result;

[0066] The preset encoding matrix is ​​fused with the mapping result to obtain the encoded signal sequence.

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

[0068] To enable the holographic converter architecture to perceive the order or spatial position of symbols in the sequence, a phase-consistent position encoding is introduced. This involves fusing a preset encoding matrix P with the mapping result. Unlike standard converters that simply add position encoding to real-valued embeddings, this method maintains a clear phase information structure by fusing the encoding matrix P with the mapping result. This ensures that the spatiotemporal relationship is encoded in a way that conforms to the physical nature of the complex-valued signal. For example, complex exponential coding or sine / cosine pair extensions can be used to ensure that the position information system is effectively represented in the complex plane.

[0069] Furthermore, the aforementioned 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 Furthermore, φ(t,d) is a linear phase term, causing equidistant phase rotations at different time steps in the complex plane; r0 represents the position encoding base amplitude, with a value greater than 0, determining the overall amplitude calibration; d represents the feature dimension index, and 0 ≤ d < d model Thus, the encoding matrix P can be obtained.

[0070] This embodiment encodes the input complex-valued signal sequence, which can completely preserve the initial amplitude and phase characteristics of each input symbol, laying the foundation for subsequent deep feature extraction.

[0071] In one embodiment, the similarity results between the query signal and the key signal of each encoded signal are calculated independently, including:

[0072] Calculate the amplitude alignment between the query signal and the key signal, and the cosine relationship with the initial phase difference. Combine the amplitude alignment and the cosine relationship to calculate the first coupling result.

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

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

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

[0076]

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

[0078] Calculate the relative phase difference between the query signal and the key signal, as described above. Based on the relative phase difference and the preset learnable α parameter, the second gating term mentioned above is calculated, namely... Specifically, α ≥ 0, where α is a learnable scalar parameter that controls the model's sensitivity to phase mismatch. When the value is small, it indicates a high phase correlation between the query signal and the key signal. In this case, the learnable parameter α is close to 1, allowing for strong attention. When the value is large, it indicates that the query signal and the key signal are phase-independent or out of phase. In this case, the learnable parameter α tends to 0, suppressing attention. In summary, learning the learnable parameter α allows the model to adaptively adjust its focus on the phase correlation scale of different data features. The learnable parameter α can be set as a learnable non-negative scalar. Using softplus(β) mapping (β is an unbounded real-valued parameter), it is necessary to ensure 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 between 0.1 and 3, which can adaptively adjust the sensitivity to phase mismatch.

[0079] Furthermore, in practical applications, the aforementioned first coupling result and second gating term can be directly heatmapped to observe the coherence strength between different tokens. It can be decomposed into The distribution can be visualized by drawing rotating arrows on the complex plane, thus showing the constructive / destructive interference paths. Furthermore, the aforementioned known structures can all be naturally generated within the above scheme, without requiring additional modifications to the holographic converter architecture.

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

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

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

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

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

[0085] in, The above similarity results, V j The above signal value, To determine the relative phase difference between the query signal and the key signal, the relative phase differences between the value signal, query signal, and key signal are first fused to obtain the above... This term means that before summing the signals in the complex-valued signal sequence, the value of each signal, signal V, is... j According to Rotation is performed on the complex plane. This rotation operation enables the subsequent summation process to realistically simulate the constructive and destructive interference of waves. This also allows the network to learn the complex phase-based dependencies and interferograms in the input sequence.

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

[0087] Obtain a pre-defined complex-valued signal training set; wherein the complex-valued signal training set carries downstream task labels;

[0088] The complex-valued signal training set is input into the initial holographic transformer architecture for training to obtain the downstream task training results. 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 backpropagated to the initial holographic transformer architecture for iterative training to generate a fully trained holographic transformer architecture. Specifically, 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 downstream task labels 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 converter architecture. In some preferred embodiments, the holographic converter architecture is trained under supervision.

[0090] First, a pre-defined complex-valued signal training set is obtained. This training set carries downstream task labels. The training set is determined based on a specific downstream task. For example, if the downstream task is to predict the target distance based on electromagnetic wave signals, then the training set includes at least the transmitted and echo signals of electromagnetic waves. Correspondingly, the downstream task labels include at least the actual target distances corresponding to the transmitted and echo signals in the training set. Similarly, in the biomedical EEG phase-amplitude coupling (PAC) detection scenario, if the downstream task is to perform classification prediction based on EEG rhythm signals to determine the presence of neurological diseases (such as schizophrenia, Parkinson's disease, Alzheimer's disease, etc.), then the training set includes at least existing EEG rhythm signals. Correspondingly, the downstream task labels include at least the correct classification results corresponding to those EEG rhythm signals, and so on.

[0091] The complex-valued signal training set is input into the initial holographic transformer architecture for training. The analysis head in the decoder obtains the training results for the downstream task corresponding to the training set. 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. Then, the results of the first loss function and the second loss function are calculated. Based on the weighted sum of the results of the first loss function and the second loss function, the overall loss function result is obtained. The gradient of the overall loss function result is backpropagated to the initial holographic transformer architecture for iterative training, generating a fully trained holographic transformer architecture.

[0092] Furthermore, the specific methods for calculating the results of the first and second loss functions are as follows: The predicted training complex-valued signal sequence is reconstructed from the training holographic self-attention calculation results obtained during the training phase (understandably, the calculation method for the training holographic self-attention calculation results during the training phase is the same as the calculation method for the holographic self-attention calculation results described above). The first loss function L recon Typically, this is the complex mean square error (MSE) or the combined error of amplitude and phase, for example:

[0093]

[0094] Where E represents the expected value of the batch, Representative of the Real Department Representative of taking the imaginary part, The square of the norm (i.e., the sum of squares in the complex domain). Through the first loss function described above, the latent representation learned by the model is forced to retain sufficient information about the original signal (including fine phase structure), thus acting as a powerful physical consistency regularizer. X is the complex-valued signal training set input to the initial holographic transformer architecture.

[0095] Then, a second loss function is determined based on the downstream task. This second loss function can be determined according to the specific type of downstream task (such as cross-entropy loss, root mean square error, etc.), and the prediction results related to the downstream task are output based on the downstream task, i.e., the aforementioned downstream task training results. The training results of the downstream task and the corresponding downstream task label y are substituted into the second loss function L. analysis In the process, we obtain the result of the second loss function, at which point we have:

[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 and second loss functions, the overall training objective is the weighted sum of these two losses, i.e., L. total =λ r L recon +λ a L analysis , where λ r and λ a These are hyperparameters used to balance reconstruction fidelity and downstream task performance. Through joint optimization based on the results of the first and second loss functions, the model is guided to learn complex-valued representations that are both physically meaningful and discriminative for specific tasks. Finally, based on the above joint optimization results, the loss function L is... total The resulting gradient is backpropagated to the initial holographic transformer architecture for iterative training, generating a fully trained holographic transformer architecture.

[0099] It should be understood that although the steps in the flowcharts of the embodiments described above are shown sequentially according to the arrows, these steps are not necessarily executed in the order indicated by the arrows. Unless explicitly stated herein, there is no strict order restriction on the execution of these steps, and they can be executed in other orders. Moreover, at least some steps in the flowcharts of the embodiments described above may include multiple steps or multiple stages. These steps or stages are not necessarily completed at the same time, but can be executed at different times. The execution order of these steps or stages is not necessarily sequential, but can be performed alternately or in turn with other steps or at least some of the steps or stages of other steps.

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

[0101] In one embodiment, such as 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 sequentially:

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

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

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

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

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

[0107] Specifically, the self-attention module divides the multidimensional space into multiple subspaces (also known as "multi-head") and independently performs the aforementioned holographic self-attention operation on each subspace, obtaining the aforementioned sub-attention calculation result on each subspace. This allows the model to pay attention to information from different representation subspaces with different interference modes in parallel. The outputs of each head are then concatenated, and the sub-attention calculation results from each subspace are fused to obtain the aforementioned holographic self-attention calculation result. Here, the multidimensional space has a dimension of d. model The dimension of a 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 result, namely, to independently execute the holographic self-attention operation described above in multiple subspaces, and to fuse the sub-attention calculation results output by each subspace to obtain the final holographic self-attention calculation result. Compared with the above method of executing the holographic self-attention operation in one space, this method can more effectively capture the diversity of data and enhance the model's ability to recognize complex data.

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

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

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

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

[0113] Specifically, after calculating the similarity result, the similarity result is input into the attention score calculation module. The attention score calculation module applies a preset scaling factor to scale the similarity result, resulting in a scaled result:

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

[0115] Furthermore, after obtaining the scaling result, it is input into the normalization module. The normalization module normalizes the scaling result based on the dimension of the key signal, obtaining the normalized result, at which point:

[0116]

[0117] The above A ij This is the result of the normalization process.

[0118] Finally, based on the relative phase difference between the value signal, query signal, and key signal, a fusion process is performed to obtain the fusion result. Based on the fusion result and the normalization result, the above holographic self-attention calculation result is obtained, at which point:

[0119] Compared to the previous method of combining similarity and fusion results, this embodiment obtains the holographic self-attention calculation result based on the fusion and normalization results, thus ensuring gradient stability. Prevent d k When the gradient explodes, it is normalized before combining the two. Normalization avoids a few maxima dominating training, which can effectively improve the convergence speed, numerical stability and performance of the model.

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

[0121] Complex-valued feedforward networks are used to activate the output of the self-attention module.

[0122] Specifically, the holographic converter architecture includes a Complex Feed-Forward Network (CFFN), which preferably consists of two complex linear transformation layers and a complex activation function between them. The complex activation function includes, but is not limited to, Concatenated Rectified Linear Unit (CReLU), Complex Rectified Linear Unit (zReLU), and Modified Rectified Linear Unit (modReLU).

[0123] Complex-valued feedforward networks and self-attention modules are connected one-to-one. Complex-valued feedforward networks include:

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

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

[0126] In summary, the aforementioned complex-valued feedforward network can be used to achieve feature extraction and information enhancement of signals.

[0127] In one embodiment, the architecture further includes a holographic transformation module, wherein the holographic transformation module includes multiple self-attention modules, multiple layer normalization modules and multiple residual connection units, and the complex-valued feedforward network and the self-attention modules are connected in a one-to-one correspondence; the layer normalization modules, the self-attention modules and the complex-valued feedforward network are connected, and the multiple complex-valued feedforward networks and the self-attention modules are connected using 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 transform module includes multiple self-attention modules, with a one-to-one connection between each self-attention module and the complex-valued feedforward network. A layer normalization module is placed before the input of each self-attention module or before the input of each complex-valued feedforward network, enabling the use of a complex-valued version of layer normalization 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 built to learn complex hierarchical complex-valued features in the data.

[0130] Each module in the aforementioned holographic converter architecture can be implemented entirely or partially through software, hardware, or a combination thereof. These modules can be embedded in hardware within or independently of the processor in a computer device, or stored in software within the memory of a computer device, so that the processor can invoke and execute the corresponding operations of each module.

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

[0132] First, the 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 complex-valued convolution operation is performed on the complex-valued signal sequence, mapping it to a higher-dimensional complex-valued latent semantic space, obtaining the mapping result. Then, the mapping result is combined with a preset encoding matrix to obtain the encoded signal sequence, ensuring that the encoded signal sequence still maintains a clear phase information structure, ensuring that the spatiotemporal relationship is encoded in a way consistent with the physical nature of the complex-valued signal. The above encoded 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 encoded signal sequence, the encoded signal sequence is input into the aforementioned holographic transformation module. The holographic transformation module includes the aforementioned multi-head self-attention module, complex-valued feedforward network, layer normalization module, and residual connection unit. The multi-head self-attention module divides the multidimensional 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. Then, the self-attention calculation results output by each subspace are concatenated to obtain the aforementioned holographic self-attention calculation result. The self-attention result is input into the complex-valued feedforward network (CFFN), which activates the holographic self-attention calculation result. Furthermore, a complex-valued version of the layer normalization module can be connected before the input of each complex-valued feedforward network or self-attention module to stabilize the training process. Residual connections can be used around multiple complex-valued feedforward networks or self-attention modules to construct a deep model to learn complex hierarchical replication features in the data. It should be further explained that the above-mentioned holographic self-attention operation includes first projecting each encoded signal in the encoded signal sequence onto the query, key, and value spaces respectively to obtain the query signal, key signal, and value signal, and then independently calculating the similarity results between the query signal and key signal of each encoded signal. Then, the similarity results are scaled using the attention score calculation module to obtain the scaled result. The scaling result is normalized in the dimension based on the key signal to obtain the normalized result. Finally, the relative phase differences between the value signal, query signal, and key signal of all encoded signals in the encoded signal sequence are fused to obtain the fusion result. Based on the fusion result and the normalization result, the above-mentioned holographic self-attention calculation result H is obtained. i Specifically, calculating H i have:

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

[0135] The holographic self-attention calculation result is input into the decoding module, which includes a reconstruction head and an analysis head. During the training phase, the first loss function result is calculated through the reconstruction head. The second loss function result L is calculated by analyzing the head. analysis The system then calculates a weighted sum of the results from the first and second loss functions through joint training. The gradient of this weighted sum is then backpropagated to the initial holographic transformer architecture for iterative training, generating a fully trained holographic transformer architecture. In the post-training usage phase, only the analysis head is needed to output the prediction results corresponding to the downstream tasks, i.e., the processing results described above. In other embodiments, the complex-valued signal sequence can also be reconstructed from the holographic self-attention calculation results using the reconstruction head during the usage phase, 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 weight initialization, and Wirtinger optimizer adaptation may also be introduced. The phase smoothing constraint is a regularization term introduced to promote the learning of locally smooth phase representations of complex signal sequences in the time or spatial domains. This penalizes excessively large phase gradients between adjacent complex signal tokens in the complex signal sequence, which helps mitigate the impact of entanglement and noise on model performance. The amplitude-phase decoupling regularization introduces an auxiliary loss to encourage the information encoded by amplitude and phase channels in intermediate layer representations to be decoupled or orthogonal to some extent, preventing the model from over-relying on a single information channel or producing trivial solutions. The dedicated complex weight initialization employs methods such as initialization based on the real and imaginary parts of an independent Gaussian distribution, or initialization methods that maintain the weight matrix satisfying specific complex norm constraints, which are crucial for the stable training of deep replication networks. The Wirtinger optimizer adaptation requires that the optimizer (such as Adam) be correctly configured to handle complex gradients calculated based on Wirtinger calculus.

[0137] Based on the same inventive concept, this application also provides a computer-readable storage medium having a computer program stored thereon, wherein the computer program, when executed by a processor, performs the following steps:

[0138] Obtain the complex-valued signal sequence to be processed, and encode the complex-valued signal sequence to obtain the encoded signal sequence;

[0139] A holographic self-attention operation is performed on the encoded signal sequence to obtain the holographic self-attention calculation result. The holographic self-attention operation includes projecting each encoded signal in the encoded signal sequence onto the query, key, and value spaces respectively to obtain the query signal, key signal, and value signal; independently calculating the similarity results between the query signal and key signal of each encoded 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 encoded signals, as well as the similarity results.

[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, data stored, data displayed, 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 understand that all or part of the processes in the above embodiments can be implemented by a computer program instructing related hardware. The computer program can be stored in a non-volatile computer-readable storage medium. When executed, the computer program can include the processes of the embodiments described above. Any references to memory, databases, 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 many forms, such as Static Random Access Memory (SRAM) or Dynamic Random Access Memory (DRAM). The databases involved in the embodiments provided in this application may include at least one type of relational database and non-relational database. Non-relational databases may include, but are not limited to, blockchain-based distributed databases. The processors involved in the embodiments provided in this application may be general-purpose processors, central processing units, graphics processing units, digital signal processors, programmable logic devices, quantum computing-based data processing logic devices, etc., and are not limited to these.

[0143] The technical features of the above embodiments can be combined in any way. For the sake of brevity, 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 embodiments described above are merely illustrative of several implementation methods of this application, and while the descriptions are specific and detailed, they should not be construed as limiting the scope of this patent application. It should be noted that those skilled in the art can make various modifications and improvements without departing from the concept of this application, and these all fall within the protection scope of this application. Therefore, the protection scope of this application should 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 includes: Obtain the complex-valued signal sequence to be processed, and encode the complex-valued signal sequence to obtain the encoded signal sequence; A holographic self-attention operation is performed on the encoded signal sequence to obtain a holographic self-attention calculation result. The holographic self-attention operation includes projecting each encoded signal in the encoded signal sequence onto a query, key, and value space 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 encoded 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 encoded signals, as well as 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 value signal sequence.

2. The method according to claim 1, characterized in that, The encoding process of the complex-valued signal sequence to obtain an encoded signal sequence includes: The complex-valued signal sequence is mapped to the complex-valued latent semantic space to obtain the mapping result; The preset encoding matrix is ​​fused with the mapping result to obtain the encoded signal sequence.

3. The method according to claim 1, characterized in that, The independent calculation of the similarity results between the query signal and the key signal of each of the encoded signals includes: Calculate the amplitude alignment degree between the query signal and the key signal, and the cosine relationship between them and the initial phase difference. Combine the amplitude alignment degree and the cosine relationship to calculate the first coupling result. Calculate the relative phase difference between the query signal and the key signal, and calculate the 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; The similarity result is obtained based on the first coupling result and the second gating term.

4. The method according to claim 1, characterized in that, The step of calculating the holographic self-attention calculation result based on the relative phase difference between the value signals, the query signals, and the key signals of all the encoded signals, and the similarity result, includes: 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 the 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, characterized in that, The method further includes: Obtain a preset complex-valued signal training set; wherein the complex-valued signal training set carries downstream task labels; The complex-valued signal training set is input into an initial holographic transformer architecture for training to obtain downstream task training results. A first loss function result and a second loss function result are calculated, and the gradients of the first and second loss function results are backpropagated to the initial holographic transformer architecture for iterative training to generate a fully trained holographic transformer architecture. Specifically, the predicted training complex-valued signal sequence is reconstructed from the training holographic self-attention calculation results obtained during 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. A corresponding second loss function is determined based on the downstream task, and the downstream task training results and the downstream task labels 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, which are connected sequentially. 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 encoded signal sequence to obtain a holographic self-attention calculation result. The holographic self-attention operation includes projecting each encoded signal in the encoded signal sequence onto a query, key, and value space 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 encoded 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 encoded 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 the 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 in each subspace to obtain the sub-attention calculation result. The sub-attention calculation results output from each subspace are then fused to obtain the holographic self-attention calculation result.

8. The architecture according to claim 6, characterized in that, 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 results based on a preset scaling factor to obtain a scaled result. The normalization module is used to normalize the scaling result based on the dimension of the key signal to obtain a normalized result. The coherent superposition unit is used to fuse the relative phase difference between the value signal, the query signal and the key signal to obtain a fusion result. The holographic self-attention calculation result is obtained based on the fusion result and the normalization result.

9. The architecture according to claim 6, characterized in that, The holographic converter architecture also includes a complex-valued feedforward network; the complex-valued feedforward network includes multiple complex-valued linear transformation layers and complex-valued activation functions, 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, multiple layer normalization modules, and multiple residual connection units, and the complex-valued feedforward network and the self-attention modules are connected in a one-to-one correspondence; the layer normalization modules, the self-attention modules, and the complex-valued feedforward network are connected, and the multiple complex-valued feedforward networks and the self-attention modules are connected using residual connections; 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, it implements the steps of the method according to any one of claims 1 to 5.

Images