Inference analysis acceleration method based on sequence neural network

Abstract

The invention provides an inference analysis acceleration method based on a sequence neural network, and the method comprises the steps: constructing a sequence neural network which is a neural network for sequence data, and comprises an intermediate layer, a plurality of intermediate layers, and an output layer; and obtaining reasoning sequence data, and performing reasoning analysis on the reasoning sequence data through the sequence neural network to obtain a reasoning analysis result. According to the technical scheme, the accuracy of reasoning analysis is improved, reasoning delay is reduced, and power consumption in the reasoning process is reduced.

Description

Technical Field

[0001] This invention relates to the field of reasoning analysis technology, and more specifically, to a method for accelerating reasoning analysis based on sequence neural networks. Background Technology

[0002] Inference analytics is applied to various complex intelligent scenarios, such as edge computing, the Internet of Things (IoT), and artificial intelligence (AI). These intelligent applications demand high response speeds for inference services; therefore, improving the speed of intelligent inference analytics has significant practical value. Currently, academic research on inference analytics has yielded fruitful results, effectively and efficiently handling various intelligent application scenarios at this stage. However, with the rise of cloud computing and IoT technologies, traditional inference analytics techniques have revealed some shortcomings. Therefore, there is an urgent need to combine current advanced sequence neural network technologies with traditional inference analytics techniques to explore new directions for inference analytics.

[0003] Many scholars have conducted scientific research on the evolution of reasoning and analytical techniques. Peng Yu-xin believes that cross-media analysis and reasoning is an active research area in computer science and a promising direction for sequence neural networks. However, existing research has not summarized the state-of-the-art cross-media analysis and reasoning methods, nor has it introduced the progress, challenges, and future directions in this field. To address these issues, he proposes methods, progress, and future directions for cross-media analysis and reasoning. His goal is not only to draw more attention to the latest developments in this field but also to provide technological insights by discussing the challenges and research directions. Tecuci Gheorghe proposes a method and system based on scientific methods. By integrating the analyst's imagination and expertise with computer knowledge and critical reasoning, it helps intelligence analysts and others to reason better when dealing with problems involving incomplete, contradictory, ambiguous, and missing information. Mehraj BHA T believes that reasoning and problem-solving skills are becoming increasingly important in making informed decisions in daily life. Current research involves background research on reasoning and problem-solving abilities. He conducted research tests on reasoning ability, problem-solving ability, and learning styles. He used analytical data. The results show that participants had different reasoning and problem-solving abilities when using different learning styles. Furthermore, students with assimilation and divergent learning styles exhibit better reasoning and problem-solving abilities. The results of this study will contribute to the literature on learning styles and cognitive abilities, providing broad implications for classroom teachers, curriculum developers, researchers, and educational planners. Berka Petr argues that with the rapid growth of various data and information on social media, sentiment analysis and reasoning are becoming increasingly popular. Determining whether a statement expresses positive, negative, or neutral sentiment towards a topic of discussion can be understood as a knowledge-based reasoning and analysis problem. Various knowledge-based reasoning techniques can be used to solve this problem. He focuses on two complementary approaches derived from the field of sequence neural networks: rule-based reasoning and case-based reasoning. He describes the basic principles, advantages, and limitations of these methods and, based on a literature review, explains how these methods can be used for sentiment analysis. The aforementioned research topics have explored reasoning analysis techniques from multiple perspectives, providing guidance and inspiration, and offering significant assistance to the research work in this project.

[0004] Today, with the rapid development of science and technology, various intelligent technologies are being applied to multiple interdisciplinary research topics. Therefore, sequence neural network technology can be applied to the research of reasoning analysis technology, which will provide more intelligent research for reasoning analysis technology. Currently, many scholars have combined sequence neural network technology with reasoning analysis technology. Ku Hao-Hsiang pointed out that big data applications have become a top priority in many fields. Many researchers are committed to improving accuracy and reducing time complexity. Therefore, he designed and proposed an ontology-based backpropagation neural network classification and reasoning strategy for big data applications and constructed a large-scale behavioral model. He carefully analyzed the classification and reasoning strategy of the backpropagation neural network, and the results showed that big data and neural network technologies can effectively build high-performance data storage, search, and retrieval environments. Ning Zhou proposed an entity coreference resolution model that integrates neural networks and global reasoning, and attempts to solve the problem of fuzzy and sparse distribution of complex entity information and reference information in text. He used a neural network model to extract entities and their antecedents and consequences from documents. Then, he combined the contextual information of sentences for global reasoning. Finally, he added the reasoning results to the neural network model to improve the accuracy of entity coreference resolution. The results show that, compared with existing models, the proposed algorithm improves the parsing performance of core words and better understands the semantic information of the text. While the research of the aforementioned scholars provides a good direction for this research, it remains at the theoretical level and lacks practicality and real-world application.

[0005] Inference analytics has been applied to many intelligent scenarios. For example, it can be used to develop intelligent web search engines and for intelligent learning of various machine languages. However, current inference and analysis methods are gradually revealing many shortcomings, such as slow data collection and aggregation, rigid inference results, and neglecting the dynamic changes in data during actual production. Therefore, inference analytics methods also need to be improved and combined with state-of-the-art cutting-edge technologies.

[0006] In summary, as the amount of computation required for intelligent reasoning increases, the challenges become increasingly apparent: First, the resource costs of computational inference are constantly rising; second, the excessive computational load of inference leads to excessively long waiting times, which significantly impacts the user experience. Therefore, improving inference performance and inference technology is particularly urgent. Summary of the Invention

[0007] In view of this, the present invention proposes a reasoning analysis acceleration method based on sequence neural networks to solve the problems existing in the prior art.

[0008] To achieve the above objectives, this invention proposes a method for accelerating inference analysis based on sequence neural networks, comprising: Construct a sequence neural network, wherein the sequence neural network is a neural network for sequence data, and the sequence neural network includes an input layer, several intermediate layers and an output layer; The inference sequence data is acquired, and the inference sequence data is analyzed using a sequence neural network to obtain the inference analysis results.

[0009] Optionally, in the sequence neural network, the intermediate layer includes a sequence processing layer, a batch normalization layer, and an activation function, wherein the sequence processing layer employs several neurons, convolutional layers, fully connected layers, LSTM layers, or a Transformer encoder.

[0010] Optionally, the reasoning sequence data is triple data or triple data sequence in a knowledge graph, used to represent the relationship between entities, and the reasoning analysis result is the entity data obtained through reasoning.

[0011] Optionally, in the sequence neural network, the output of the batch normalization layer is:

[0012]

[0013]

[0014] Where z represents the normalized output value, μ represents the mean of all output data y in the current batch, V represents the variance of all output data in the current batch, γ represents the scaling parameter, and β represents the offset parameter. Let n be a positive number, i be the channel number, and x be the channel ID. i Let b represent the i-th input data, b represent the bias term, and w represent the bias term. i The weights of the convolution kernel, W i B represents the normalized weight, and B represents the normalized offset.

[0015] Optionally, the intermediate layer has residual connections between the input and the output.

[0016] Optionally, the sequence neural network further includes an energy prediction module and an early exit classifier. The energy prediction module is located at the output of the intermediate layer, and the early exit classifier is located at the output of the energy prediction module. The energy prediction module evaluates the energy of the output data of the intermediate layer to obtain an energy value. A threshold judgment is made on the energy value, and the output of the intermediate layer is selected to be input to the next intermediate layer or the early exit classifier based on the energy value. Direct inference analysis is performed through the early exit classifier to obtain the inference analysis result.

[0017] Optionally, in the sequence neural network, a hybrid inference layer is also provided between the intermediate layers. The hybrid inference layer converts the output data of the intermediate layers into symbolic logic assertions, performs forward chain inference on the symbolic logic assertions to obtain new symbolic conclusions, re-encodes the new symbolic conclusions and symbolic logic assertions, and fuses the encoding results with the output data of the intermediate layers as the input data of the next intermediate layer.

[0018] Optionally, the sequence neural network is further configured with several branch structures, each including different branches for processing multimodal data. Each branch is a sequence neural network structure. Each branch fuses the output data of the corresponding intermediate layer and outputs it to the next intermediate layer. Between output layers, the output results of different branches are fused.

[0019] On the other hand, the present invention provides an inference analysis acceleration system based on sequence neural networks for performing the above-described method.

[0020] Compared with the prior art, the beneficial effects of the present invention are as follows: This invention proposes an evolutionary reasoning analysis technique based on sequence neural networks, aiming to combine traditional reasoning analysis techniques with sequence neural network technology. The proposed algorithm is an accelerated reasoning analysis algorithm based on sequence neural networks. This algorithm can utilize each layer of neurons in the neural network to analyze and process large amounts of data. This algorithm can improve the accuracy of reasoning analysis, reduce inference latency, and lower power consumption during the inference process. Through testing and investigation of the sequence neural network-based reasoning analysis technique, the results show that the inference latency is 648ms, the inference accuracy is 97.5%, and the power consumption is 210W. The test results are superior to traditional reasoning analysis techniques. These test results strongly demonstrate the practical value of research on reasoning analysis based on sequence neural networks. This research result strongly indicates that sequence neural network technology can be well applied to current reasoning analysis techniques, pointing to an advanced development path for future evolutionary reasoning analysis techniques. Attached Figure Description

[0021] Various other advantages and benefits will become apparent to those skilled in the art upon reading the following detailed description of preferred embodiments. The accompanying drawings are for illustrative purposes only and are not intended to limit the invention. In the drawings: Figure 1 This is the intelligent feature map of the reasoning evolution technology in the embodiments of the present invention; Figure 2 This is a diagram illustrating the main pattern of reasoning analysis in this embodiment of the invention. Figure 3These are related technical diagrams for reasoning analysis in embodiments of the present invention; Figure 4 This refers to the ratio of computational cost and parameter quantity for each neural network layer in this embodiment of the invention. Figure 5 This is a comparison of the performance indicators of different reasoning analysis techniques in the embodiments of the present invention; Figure 6 This is a schematic diagram of the method flow in an embodiment of the present invention. Detailed Implementation

[0022] Exemplary embodiments of the present disclosure will now be described in more detail with reference to the accompanying drawings. While exemplary embodiments of the present disclosure are shown in the drawings, it should be understood that the present disclosure may be implemented in various forms and should not be limited to the embodiments set forth herein. Rather, these embodiments are provided to enable a more thorough understanding of the present disclosure and to fully convey the scope of the disclosure to those skilled in the art. It should be noted that, unless otherwise specified, the embodiments and features described herein can be combined with each other. The present invention will now be described in detail with reference to the accompanying drawings and embodiments.

[0023] Research on reasoning evolution analysis techniques based on sequence neural networks is currently a focus of social attention. With the acceleration of intelligentization, various intelligent devices are being applied to various production and life scenarios. However, with the increase in intelligent computing power, the dilemmas are becoming increasingly apparent: firstly, the resource costs of computational reasoning are constantly rising; secondly, the excessive computational load leads to long waiting times, which greatly affects the user experience. Therefore, improving reasoning performance and reasoning technology is particularly urgent. Fortunately, fuzzy logic and sequence neural network technologies perform exceptionally well in the field of reasoning evolution. Therefore, combining traditional reasoning evolution analysis techniques with sequence neural network technology, and using sequence neural networks to optimize reasoning analysis techniques, is an excellent idea. Currently, reasoning technology suffers from many problems, making current reasoning and analysis techniques unable to adapt to today's intelligent information scenarios.

[0024] This invention proposes an evolutionary reasoning analysis technique based on sequence neural networks, aiming to combine traditional reasoning analysis techniques with sequence neural network technology. The proposed algorithm is an accelerated reasoning analysis algorithm based on sequence neural networks. This algorithm can utilize each neuron layer, or data processing layer, of the neural network to analyze and process large amounts of data. This algorithm can improve the accuracy of reasoning analysis, reduce inference latency, and lower power consumption during the inference process. Through testing and investigation of the sequence neural network-based reasoning analysis technique, the results show that the inference latency is 648ms, the inference accuracy is 97.5%, and the power consumption is 210W. The test results are superior to traditional reasoning analysis techniques. These test results strongly demonstrate the practical value of research on reasoning analysis based on sequence neural networks. This research result strongly indicates that sequence neural network technology can be well applied to current reasoning analysis techniques, pointing to an advanced development path for future evolutionary reasoning analysis techniques.

[0025] This section provides an introduction to reasoning techniques: Definition of inference evolution technology: Currently, science and technology are increasingly interdisciplinary. While information technology has transformed human production and lifestyles, the amount of information in human production and life is increasing exponentially, thus placing higher demands on information reasoning technology. In various human activities, more intelligent and rapid information reasoning and analysis technologies are needed to help people process massive amounts of information. Artificial intelligence has greatly facilitated the learning and application of knowledge. Inference evolution technology can integrate intelligent learning and evolutionary strategies into reasoning algorithms, forming a hybrid intelligent evolutionary reasoning algorithm with knowledge learning and continuous evolution, thereby continuously improving the performance of reasoning analysis algorithms. Inference evolution technology differs from other intelligent algorithm technologies in that it possesses many desirable characteristics of natural biology. This algorithm can mimic the evolutionary rules of intelligent animals, continuously iterating and evolving in cycles, and constantly adapting to various computational reasoning scenarios. The intelligence of inference evolution technology mainly includes: self-learning, self-iteration, and adaptability. Specific intelligent characteristics are as follows: (1) Self-learning: Learning ability is a unique intelligent behavior of intelligent animals. Reasoning evolution technology can continuously improve reasoning efficiency by selecting and exchanging operators, which helps to improve the fluency of reasoning and knowledge learning.

[0026] (2) Self-iteration: No additional information from the external environment is required during the inference process. The performance of the inference algorithm can be gradually optimized and enhanced by performing special iterative optimization operations of the evolutionary algorithm.

[0027] (3) Adaptability: Adaptability is a unique ability of intelligent organisms. In each iteration of reasoning evolutionary computation, it continuously adapts to the working environment and can adapt to some special problems. It can deepen the understanding of the environment and the problem, improve the reasoning speed and expand the reasoning scope.

[0028] In summary, as an intelligent technology, reasoning evolution technology can continuously iterate and improve its reasoning performance during use. The intelligent characteristics of reasoning evolution technology include... Figure 1 As shown.

[0029] Definition and main patterns of reasoning techniques: Processing and understanding human knowledge is the ultimate goal of artificial intelligence research. Against the backdrop of an explosion of human knowledge, how to effectively store, manage, and analyze this massive amount of data is an urgent problem. At this juncture, knowledge graphs present tremendous opportunities for development. Knowledge graphs not only contain relationships between data but also indicate the semantic relationships between them. They are gradually becoming the foundation of web data. The emergence of knowledge graphs has ushered in a more intelligent era of the Internet of Things (IoT).

[0030] In knowledge graphs, when facing massive data processing challenges, the primary focus is on analyzing triplet relationships within the data: head entity, relation, and tail entity. This naturally leads to the advancement of intelligent reasoning and analytics techniques. Their main objective is to automatically estimate the latent relationships between related entity pairs and predict entities using intelligent algorithms such as machine learning and neural networks. For a long time, reasoning analytics has not only been an important theoretical application in statistical analysis and artificial intelligence learning but also a significant driving force in the field of knowledge graphs. Currently, research on reasoning analytics has achieved significant progress, with the main modes of reasoning analytics as follows: (1) Simple Reasoning: The simple reasoning model is similar to information prediction technology. It can infer the relationship between two given entities based on existing entity information and relationships on the knowledge graph. The difficulty lies in understanding the meaning of the information and relationships of existing objects. If there are head entities and tail entities, the understanding and reasoning process of simple reasoning can be completed, and information matching and reasoning can be completed. In this way, triples can be created as much as possible. However, on a further reasoning test dataset, the reasoning performance of this method is significantly better than other methods. Simple reasoning is the simplest and easiest-to-use reasoning model among geometry-based reasoning models.

[0031] (2) Inductive Reasoning: Inductive reasoning shares some similarities with simple reasoning. Both reasoning models involve reasoning processes based on information prediction. However, there are also differences between them. Because the structures of data information sets are different, it is difficult to disseminate and summarize knowledge information into the data information sets. When matching data objects on a knowledge graph, the core of the inductive reasoning model is learning the relationships between entities and their internal semantic structures.

[0032] (3) Multi-step reasoning: When reasoning about complex problems, the problem often consists of multiple steps. For example, query-based reasoning problems can be solved using simple reasoning or inductive reasoning models, but they suffer from structural diversity, non-logical operations, and other issues, resulting in extremely high computational complexity for reasoning and analysis. Furthermore, as the reasoning and analysis process progresses, the number of entities increases exponentially from the initial nodes. To overcome this problem, multi-step reasoning models have emerged. In multi-step reasoning models, the structure and data of the reasoning problem are first input, and then the final characteristics of the problem are obtained through multi-stage reasoning logic operations. These characteristics are then used for further comparative analysis of the problem to obtain the final reasoning answer.

[0033] The main patterns of reasoning analysis are as follows Figure 2 As shown.

[0034] Related techniques for reasoning analysis: In recent years, considerable research achievements have been made in intelligent reasoning and analysis technologies. Currently, the amount of network information data is growing exponentially. Data-driven reasoning and analysis technologies have become the mainstream of current research. Reasoning and analysis technologies can be divided into: logic rule-based reasoning, distributed representation-based reasoning, and neural network-based reasoning. This chapter details the basic methods of current reasoning and analysis technologies.

[0035] (1) Reasoning techniques based on logical rules Logical rule-based reasoning can realize the information reasoning process through manually constructed rule models, and interpret logical inferences about events and information through semantic logic and structured information. This reasoning method matches single-order logical rules with different probability models and establishes a logical network for relational reasoning, thereby obtaining the basis for reasoning. Because most logical reasoning rules originate from real life, the results of this reasoning model are essentially similar to manual reasoning. In addition, some scholars use knowledge ontology content on knowledge graphs to estimate the attributes of data.

[0036] (2) Inference technology based on distributed representation In the past, reasoning and analysis of unknown information largely relied on logical rules and lacked the random walk behavior found in parallel corpora. In recent years, rule-based reasoning has become increasingly reliant on logical rules, and the diversity and effectiveness of these rules have severely impacted model effectiveness. Distributed representation-based reasoning analysis techniques can obtain distributed representations of entities, relations, and attributes in a dataset by mapping them to a vector space, which effectively improves reasoning clarity. Therefore, distributed representation-based reasoning techniques have achieved significant success in many fields, including data reasoning, attracting considerable research attention and leading to the emergence of numerous embedding models. These models obtain distributed views by mapping entities, relations, and attributes on a knowledge graph to a consistent vector space.

[0037] (3) Neural network-based inference techniques With the advent of the deep learning era, intelligent inference technology has introduced neural network models such as recurrent neural networks, convolutional neural networks, and sequence neural networks. Neural network-based inference technology is currently very popular. Its core is to understand the best representation of each node and assign further label prediction tasks based on that node. At this point, effective nodes are learned by transmitting neural information. First, each node has initial node features. Different levels of graphs can be used to integrate network information, node neighbor information, and node information, and update the node view.

[0038] The advantage of neural networks lies in their ability to learn more efficient node representations. Furthermore, the learned nodes of multi-layer nonlinear graph networks can be used to represent information, thus possessing strong reasoning and analytical capabilities. However, a drawback of neural networks is that each node is independent when making predictions based on its trained representation. It does not simulate the relationships between nodes and lacks some information connectivity. The practical applications of neural networks in reasoning analysis are mainly as follows: Neural network tensor models can use bilinear tensor layers to combine relationships and can be used to describe complex semantic relationships between entities. Furthermore, bidirectional neural network algorithms can further capture path sequence relationships and improve the model's reasoning ability. Finally, introducing convolutional neural networks into the reasoning analysis process and directly capturing the interaction factors of triples through sliding modes can also greatly improve reasoning efficiency. Among these, related techniques for reasoning analysis include... Figure 3 As shown.

[0039] Problems faced by reasoning analysis: Early research on reasoning primarily focused on logic and knowledge-intensive domains. Almost all reasoning was based on existing knowledge logic, emphasizing how to derive correct conclusions from known sentences and predicates. However, with the explosive growth of data on the internet, traditional symbolic reasoning methods can no longer meet the demands of knowledge computing in the era of big data. Against this backdrop, reasoning methods based on knowledge graph variation representations have gradually become a major research direction. This method efficiently computes semantic relationships between entities and relations in low-dimensional space, effectively solving the problems of data sparsity and computational efficiency, and significantly improving the reasoning performance of knowledge graphs.

[0040] However, reasoning analysis still faces difficulties and challenges. First, in reasoning analysis, there are many complex relationships between entities. In these aspects, a head entity may have multiple tail entities, forming incorrect triples. This leads to the inefficiency of existing reasoning models when handling complex relationships. Second, existing intelligent reasoning models often require a large number of entity and relation samples to accurately study the relevant semantic information. Furthermore, in most knowledge graphs, due to the limited number of relationships associated with them, it is difficult to study the semantic features of certain structures, which to some extent limits the performance of intelligent reasoning models. Finally, most knowledge graphs cannot be accurately constructed at present, leading to ambiguity in many sentences and relational words. Therefore, in the face of complex and ambiguous reasoning, it is crucial to complete reasoning quickly and accurately.

[0041] Sequence neural network technology: Fuzzy logic technology: Fuzzy logic is a reasoning pattern similar to human reasoning and also resembles the behavior of artificial intelligence. Fuzzy logic techniques encompass all intermediate possibilities between certainty and negation. It applies a specific input probability level to generate a specific choice output. In fact, there are multiple implementation methods that can operate in a variety of environments. It can typically be implemented in controllers, network systems, and other systems of varying sizes and functions, and can be simulated through hardware and software.

[0042] Fuzzy logic offers simple reasoning similar to human reasoning. Using this logic has several advantages, such as the relatively simple structure of fuzzy logic systems, their wide applicability and powerful performance, and their ability to handle many uncertainties in work. These are the strengths of fuzzy logic technology. However, it also has some drawbacks. Because the calculation method is uncertain, the results are often hypothetical. Furthermore, the calculation and analysis of fuzzy logic require extensive debugging with specialized hardware, resulting in high costs.

[0043] With the increasing demand for intelligent computing, simulation computing technology based on fuzzy logic has emerged. It delves into intelligent computing analysis from various aspects, effectively improving the speed of intelligent computing and reducing its resource consumption. It plays a crucial role in the entire process of intelligentizing computational simulation.

[0044] Overview of Sequence Neural Networks: As a commonly used neural network model, a sequence neural network is a special structure that maps input neurons to output neurons. Users can build sequence neural networks by stacking simple data into single-layer structures. Each layer of a neural network is a formula that transforms input data into output data. The process of computing the output of a layer using a dataset is called the maximum one-pass propagation. Similarly, to compute the direct propagation of the entire sequence neural network, the direct propagation of each layer can be computed sequentially from the input vector.

[0045] In recent years, with advancements in sequence neural network algorithms and optimizations in deep learning technology, the applications of sequence neural networks have become possible. They are primarily used in time series forecasting, mathematical function approximation, financial trend prediction, and robot control. Sequence neural networks are used to further improve the accuracy of intelligent computing and effectively promote its development. Furthermore, the application areas of sequence neural networks are still being explored and optimized, and they will be further applied to more intelligent fields.

[0046] Inference analysis acceleration algorithm based on sequence neural networks: In the computation of sequential convolutional neural networks, batch normalization after standard convolutional layers and depth levels is a common approach. Large-scale normalization overcomes the stepwise learning difficulties of sequential neural networks. Furthermore, to address the issues of slow gradient convergence and data computation explosion, normalization is performed sequentially using groups during training. Generally, adding batch normalization layers to sequential neural networks can improve self-learning speed and optimize the accuracy of the computational model.

[0047] In a sequence neural network, the output of the previous network layer is used as input, and a single output data is calculated: (1) Where z represents the normalized output value, i.e., the output of the current layer. y represents the output data of the previous network layer (i.e., the input of the current layer). μ represents the mean of all output data y in the current batch. V represents the variance of all output data y in the current batch. γ represents the scaling parameter, a learnable parameter used to adjust the normalized scale. β represents the offset parameter, a learnable parameter used to adjust the normalized offset. ε represents a very small positive number (e.g., 10). -5 (), used to prevent the denominator from being zero and to ensure numerical stability.

[0048] In formula (1), y is the output data of the previous network layer, is the average value of all output data, is the variance of these output data, are two important parameters required for the normalization layer, and is an additional parameter for calculation.

[0049] Integrating the normalization layer into the network layer is relatively simple. Based on the sequence calculation method, y in formula (1) is expanded as follows: (2) Where wi represents the weight of the convolution kernel on the i-th input channel. xi represents the i-th input feature map or input data. b represents the bias term. In the above formula, wi is the weight of the convolution kernel, and b is the bias. Then, substituting these values ​​into formula (1): (3) in, It represents an extremely small positive number, as with ε above.

[0050] By rearranging formula (3), we get: (4) For formula (4), and can be expanded to obtain: (5) (6) Then formula (4) becomes: (7) The mean and variance values ​​of the grouping layers used in inference are also the sample and variance values ​​of the entire training dataset estimated using the moving average during network learning, and are therefore determined during inference. The convolutional kernel parameters of the convolutional layers can be combined with the grouping normalization layers, which can reduce the amount of computation and access during inference, thereby alleviating the resource consumption required for computation.

[0051] like Figure 6 As shown, the sequence neural network described above comprises an input layer, several intermediate layers, and an output layer. The intermediate layers include a sequence processing layer, a batch normalization layer, and an activation function. The sequence processing layer employs several neurons, convolutional layers, fully connected layers, LSTM layers, or a Transformer encoder. The input data for the input layer consists of triplet data or triplet data sequences. The entities and relations within the triplet data are converted into entity embeddings and relation embeddings using existing transformers or transformation rules and then input into the input layer. Entity embeddings represent each entity (e.g., "Beijing", "China") as a fixed-length vector. Relation embeddings represent each relation (e.g., "the capital is located in") as a vector. The output layer outputs the tail entity relations or the inference results (entities) of the above sequence under different probabilities. The triplet data sequence includes a series of inference chains between entities and relations.

[0052] In the above sequence neural network, as an example, the intermediate layer can be three layers, where the first two layers are sequence processing layers using Transformer encoders or convolutional layers, and the last layer is a fully connected layer. The specific structure (number of layers and type) of the above structure is constructed according to the reasoning capability and actual needs.

[0053] Meanwhile, residual connections are also set in the sequence processing layer, normalization layer and activation function of the intermediate layer to preserve different original features as much as possible.

[0054] Meanwhile, for better inference and improved inference efficiency, in the above content, the present invention also provides a dynamic network pruning method based on the energy of the inference path. In this method, mainly based on the above sequence neural network, an energy prediction module and multiple early exit classifiers are introduced. The energy prediction module is a lightweight fully connected layer, located after a specific intermediate layer of the backbone sequence neural network (for example, after the first and second layers). Each of the early exit classifiers consists of a fully connected layer and a batch normalization layer, and its output dimension is the same as that of the final output layer. <0>< / <0>< /

[0055] In the above content, the input data in the input layer flows through the sequence neural network to the layer where the first energy prediction module is located (for example, the intermediate layer of the first layer). This module outputs a scalar energy value E based on the activation values of the current layer, and this value represents the remaining computational complexity required to complete the inference of the current sample. The energy value E is compared with a preset and trainable threshold Ti. If E < Ti, it is considered that the sample is already easy to infer, and it is immediately guided to the corresponding early exit classifier and the final result is output, skipping all subsequent network layers. If E >= Ti, the sample continues to flow to deeper network layers. This process can be repeated at subsequent energy prediction points until the sample reaches the end of the network. During training, the total loss function is the weighted sum of the losses at each exit and the loss at the final exit, so as to jointly optimize all classifiers and energy thresholds. By setting different thresholds Ti, precise control of the inference speed and accuracy can be achieved. It can adaptively allocate computing resources, significantly reducing the average inference latency and system energy consumption of simple samples. It can work perfectly in coordination with the batch normalization acceleration technology, enjoying the acceleration benefits brought by computational fusion at each exit, and is particularly suitable for edge intelligence inference scenarios with extremely high requirements for real-time performance. <0>< / <0>< /

[0056] To enhance the model's logical reasoning capabilities and interpretability, a symbolic-neural hybrid reasoning layer is inserted into the intermediate layers of the sequence neural network (e.g., between two modules). This hybrid reasoning layer comprises an interpreter, a symbolic reasoning engine, and an embedder. The output vectors of the aforementioned sequence neural network layer are fed into the interpreter. This interpreter is a trained, lightweight neural network tasked with decoding high-dimensional vectors into probabilistic symbolic logical assertions (e.g., outputting the predicate LocatedIn(Beijing, China) with a certain confidence level). The generated symbolic assertions are fed into the symbolic reasoning engine. Based on predefined or learned logical rules, or through relevant relational queries using existing relational databases (e.g., generating IsCapitalCity(X) from LocatedIn(X, Y) and IsCapital(Y)), the engine performs rapid forward chain reasoning to derive new symbolic conclusions. The newly derived symbolic conclusions, along with the original assertions, are re-encoded into vector form by the embedder (another lightweight neural network). The vector is fused with the original neural flow (the original output of the current intermediate layer) through a gating mechanism (such as a gated recurrent unit, GRU) before being input into subsequent network layers. The interpreter and embedder are trained end-to-end with the main network. The rules of the symbolic reasoning engine can be predefined or optimized through reinforcement learning. Embedding interpretable symbolic reasoning into a learnable neural network breaks the "black box" limitation of neural networks. It gives the model the ability to perform explicit logical deduction, enabling it to handle complex, long-chain reasoning tasks that are difficult for pure neural networks to solve, thus achieving higher accuracy in applications such as knowledge graph completion and question answering systems. It provides clear reasoning path tracing, allowing users to see which logical rules led to the final decision.

[0057] Single sequence neural networks often struggle to deeply capture the complex relationships between modalities when handling complex reasoning tasks that integrate multiple modalities such as text, knowledge graphs, and time-series data. This embodiment constructs a multi-branch parallel, deeply interactive sequence neural network architecture to achieve dynamic information complementarity between modalities at various stages of the reasoning process, thereby significantly improving the joint reasoning ability for complex problems. This embodiment constructs a multi-branch sequence neural network that starts in parallel from the input layer and deeply interacts in multiple intermediate layers. This architecture itself is an advanced form of sequence neural network, consisting of multiple parallel sub-sequence neural networks (branches), each branch processing data of one modality. Each branch starts at the input layer and extends throughout the entire network depth. The different branches are as follows: Branch A (text modality): receives a sequence of text embedding vectors. Branch B (knowledge modality): receives a sequence of knowledge graph entity / relation vectors. Branch C (time-series modality): receives a sequence of time-series signals such as sensor data. After multiple intermediate layers (e.g., layers 1, 2, and 3) of each branch, a cross-modal time-series attention interaction module is set up.

[0058] The data processing flow is as follows: Step 1: Modality-Specific Serialization Processing Input: The raw data of different modalities are preprocessed and embedded by their respective layers to be transformed into corresponding serialized vectors. Text branch input: H_text^(0) = [w1, w2, ..., wn]; Knowledge branch input: H_knowledge^(0) = [e1, r1, e2, ...]; Time branch input: H_time^(0) = [s1, s2, ..., st]. w represents text data, e and r represent entity and relation data of triples, and s represents time-series data with time-series features.

[0059] Each branch acts as an independent sequence neural network (which can be a Transformer or LSTM, etc.) to extract modality-specific features from the input sequence.

[0060] H_text^(l) = TransformerLayer_text(H_text^(l-1)); H_knowledge^(l)=TransformerLayer_knowledge(H_knowledge^(l-1)) H_time^(l) = TransformerLayer_time(H_time^(l-1)) The output of each branch in the l-th layer of the sequence neural network is the sequence of hidden states H_modality^(l) of that modality at the current abstraction level.

[0061] Step 2: Cross-modal temporal attention interaction Input: When all branches are computed in parallel to the preset interaction layer l, the hidden states H_text^(l), H_knowledge^(l), and H_time^(l) of each branch are fed into the cross-modal temporal attention interaction module. Query-key-value pair construction: One branch is used as the query party, and all branches (including itself) are used as key-value pairs. For example, using the text branch as the query party: Query = H_text^(l) Keys = Values ​​= Concat(H_text^(l),H_knowledge^(l), H_time^(l)) Cross-attention calculation: Calculates the attention distribution between the query and the fused keys in the text branch. This allows each word or fragment in the text branch to dynamically focus on the most relevant information across all modal sequences (which could be an entity, a relation, or a point-in-time feature).

[0062] Attention_weights = Softmax( Query • Keys^T / √d_k ) Context_text = Attention_weights • Values Gated fusion update: The obtained context vector Context_text is fused with the original state H_text^(l) of the text branch to generate the updated state.

[0063] gate = σ( W_fuse • [H_text^(l); Context_text] + b_fuse ); H_text^(l)_new = gate * H_text^(l) + (1 - gate) * Context_text; Where σ represents the activation function, W represents the weights, and gate represents the state.

[0064] The above process is executed once for each branch, meaning each branch takes turns acting as the queryer, retrieving information from other modalities. Therefore, at the end of the interaction layer, the hidden state of each branch has been enhanced and corrected by information from other modalities.

[0065] The updated branch states H_text^(l)_new, H_knowledge^(l)_new, and H_time^(l)_new are used as the inputs to the (l+1)th layer of their respective branches. Each branch continues to perform sequence computation independently until the next interaction layer or the final output layer.

[0066] At the end of the network, the final hidden states of all branches are concatenated or weighted averaged, and then passed through a fully connected classifier / regressor to output the final joint inference result.

[0067] This embodiment constructs a multi-branch sequence neural network that dynamically interacts across multiple depths. Unlike simple early or late fusion, this architecture allows information from different modalities to interact at multiple abstraction levels, simulating the human cognitive process of synthesizing multiple information sources for progressively deeper reasoning. Through an attention mechanism, each modality can dynamically retrieve relevant information from other modalities at the moment when its information is most ambiguous or most in need of corroboration, greatly enhancing the model's robustness. The number of branches and interaction positions can be flexibly adjusted according to task requirements, exhibiting strong versatility.

[0068] Implementation and testing of inference techniques based on sequence neural networks: Test of an accelerated inference analysis algorithm based on sequence neural networks After proposing the inference analysis acceleration algorithm based on sequence neural networks, it is necessary not only to rely on theory but also to test the algorithm's computational performance in practical applications. In this experiment, a pairwise comparison method was used to test the inference analysis of the same knowledge graph using traditional inference analysis techniques and the sequence neural network inference analysis acceleration algorithm.

[0069] To eliminate interference from external factors, this test selected the same processing equipment and performed inference analysis on the same knowledge graph. The following performance metrics were tested. The test results are shown in Table 1, which compares the performance of different inference analysis techniques.

[0070] Table 1 ; (1) True Positive: The ratio of the number of samples predicted to be positive to the number of actual positive samples.

[0071] (2) Loss: Error is the proportion of errors in the prediction results. The smaller the error, the more stable the calculation.

[0072] (3) Accuracy: Accuracy is the ratio of the number of correctly predicted samples to the total number of predicted samples.

[0073] (4) False Positive: The false positive rate is the ratio of the number of prediction errors to the total number of predictions.

[0074] The test results in Table 1 show that the accelerated inference analysis algorithm based on sequence neural networks has a true positive rate of 94.8% and an accuracy rate of 95.2%, both higher than the traditional inference analysis algorithm. Furthermore, the accelerated inference analysis algorithm based on sequence neural networks has a lower error rate and false detection rate than the traditional inference analysis algorithm. This means that the accelerated inference analysis algorithm based on sequence neural networks has better inference analysis accuracy and a lower error rate. From the above results, it can be concluded that the accelerated inference analysis algorithm based on sequence neural networks is superior to the traditional inference analysis algorithm.

[0075] Reasoning and analysis technology experiment: (1) Research direction The purpose of this investigation is to test the effectiveness of the inference analysis acceleration algorithm based on sequence neural networks in actual inference analysis and to explore its performance in real-world environments.

[0076] (2) Survey content This investigation focuses on two main aspects. First, it examines the proportion of computation and parameters across each layer in the sequence neural network-based accelerated inference analysis algorithm. Second, it investigates and compares the sequence neural network-based accelerated inference analysis algorithm with traditional inference analysis techniques. The aim is to investigate the degree of optimization of this algorithm relative to traditional inference analysis techniques. Through these two aspects of investigation and research, a comprehensive understanding of the theoretical and practical significance of research on sequence neural network-based inference analysis can be achieved.

[0077] (3) Survey methods This survey employs a comparative survey model. Scientific and logical experimental investigations are used to ensure the scientific rigor and validity of the survey.

[0078] (4) Survey results During the investigation, the computational load and parameter ratio of each neural network layer were first examined during the execution of the inference analysis acceleration algorithm based on sequence neural networks. The results are as follows: Figure 4 As shown.

[0079] from Figure 4It can be seen that during the operation of the inference analysis acceleration algorithm based on sequence neural networks, the computational load of the first-level network layer accounts for 34.25%, the second-level network layer for 32.25%, and the third-level network layer for 33.50%. Furthermore, the parameter proportions of each neural network layer are 35.24%, 32.66%, and 32.10%, respectively. This shows that the computational load and parameter proportions of each neural network layer are evenly distributed. This indicates that the resource utilization of each neural network layer is relatively balanced during the operation of the inference analysis acceleration algorithm. This demonstrates that the inference analysis acceleration algorithm based on sequence neural networks can fully utilize neural network resources and improve network resource utilization efficiency.

[0080] Furthermore, it is necessary to investigate the actual reasoning performance of the accelerated reasoning analysis algorithm based on sequence neural networks compared to traditional reasoning analysis techniques. The selected reasoning analysis performance metrics are as follows, and the test results are as follows: Figure 5 As shown.

[0081] The performance metrics for these inference analyses are explained below: Inference latency: Latency typically refers to the time spent throughout the entire inference process, including data preprocessing, model inference, data transmission, and inference feedback. Inference latency has a significant impact on the user experience of intelligent services. In intelligent inference analytics applications, some low-latency intelligent applications, such as augmented reality (AR), virtual reality (VR), and autonomous driving technologies, typically have very strict latency requirements.

[0082] Inference accuracy: Inference accuracy refers to the proportion of accurately predicted input samples out of the total input samples, representing the performance of inference analysis techniques. Besides the model's inherent inference ability, the speed of the input model also affects accuracy.

[0083] Energy consumption: Energy consumption for inference computing and data analysis is critical. For edge intelligence applications, energy efficiency is paramount. Low power consumption not only saves resources but also reduces heat generation in computing devices. Figure 5 middle, Figure 5 (a) in the diagram is a comparison chart of inference delays. Figure 5 (b) in the figure is a comparison chart of inference accuracy. Figure 5 (c) in the diagram is an energy consumption comparison chart.

[0084] according to Figure 5The inference latency of traditional inference analytics and the inference analytics acceleration technology based on sequence neural networks are 851ms and 648ms, respectively, with inference accuracies of 92.6% and 97.5%, and energy consumption of 240W and 210W, respectively. The results show that the inference analytics acceleration mode based on sequence neural networks outperforms the traditional inference analytics technology in terms of inference latency, inference accuracy, and energy consumption. The new inference analytics acceleration mode performs better, performs inference faster, improves the experience of intelligent services, and has higher inference accuracy. Furthermore, its lower power consumption effectively extends the lifespan of inference computing devices.

[0085] Based on the above experimental results, it can be effectively demonstrated that the inference parameters of the inference analysis acceleration algorithm have been improved. This also indicates that the research on evolutionary inference analysis technology based on sequence neural networks is real and feasible. The research content can enrich the application fields of neural networks. In conclusion, the research on evolutionary inference analysis technology based on sequence neural networks has significant practical implications.

[0086] in conclusion: Intelligent reasoning and analysis technology based on sequence neural networks not only improves the accuracy of reasoning but also reduces the latency of reasoning and analysis. It demonstrates excellent performance in real-world testing. This reasoning and analysis technology fully utilizes the computing resources of computing devices by employing excellent sequence neural network algorithms, reducing the overall computational load and significantly improving the efficiency of reasoning computation. However, despite its effective reasoning and analysis capabilities, a drawback of this technology is the need for powerful processing capabilities to transmit and process complex and massive amounts of reasoning information and data. Therefore, powerful processing equipment is essential to ensure the operation of sequence neural network reasoning and analysis technology, which is its limitation. If hardware conditions are not guaranteed, the smooth operation of this technology will be difficult to ensure. However, society will continue to progress, and science and technology will continue to innovate. In the near future, hardware will no longer be a constraint on science and technology. Therefore, the future direction is to combine traditional reasoning and analysis techniques with sequence neural networks to achieve reasoning and analysis technology based on hybrid sequence neural networks, further improving the computational performance of reasoning and analysis and opening up a new path for intelligent reasoning and analysis technology.

[0087] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention and not to limit it. Although the present invention has been described in detail with reference to the above embodiments, those skilled in the art should understand that modifications or equivalent substitutions can still be made to the specific implementation of the present invention. Any modifications or equivalent substitutions that do not depart from the spirit and scope of the present invention should be covered within the scope of protection of the claims of the present invention.

Claims

1. A method for accelerating inference analysis based on sequence neural networks, characterized in that, include: Construct a sequence neural network, wherein the sequence neural network is a neural network for sequence data, and the sequence neural network includes an input layer, several intermediate layers and an output layer; The inference sequence data is acquired, and the inference sequence data is analyzed using a sequence neural network to obtain the inference analysis results.

2. The method according to claim 1, characterized in that, In the sequence neural network, the intermediate layer includes a sequence processing layer, a batch normalization layer, and an activation function, wherein the sequence processing layer employs several neurons, convolutional layers, fully connected layers, LSTM layers, or Transformer encoders.

3. The method according to claim 1, characterized in that, The reasoning sequence data is triple data or triple data sequence in the knowledge graph, used to represent the relationship between entities, and the reasoning analysis result is the entity data obtained through reasoning.

4. The method according to claim 1, characterized in that, In the sequence neural network, the output of the batch normalization layer is: ， ， ， Where z represents the normalized output value, μ represents the mean of all output data y in the current batch, V represents the variance of all output data in the current batch, γ represents the scaling parameter, and β represents the offset parameter. Let n be a positive number, i be the channel number, and x be the channel ID. i Let b represent the i-th input data, b represent the bias term, and w represent the bias term. i The weights of the convolution kernel, W i B represents the normalized weight, and B represents the normalized offset.

5. The method according to claim 1, characterized in that, The intermediate layer has residual connections between the input and the output.

6. The method according to claim 1, characterized in that, The sequence neural network also includes an energy prediction module and an early exit classifier. The energy prediction module is located at the output of the intermediate layer, and the early exit classifier is located at the output of the energy prediction module. The energy prediction module evaluates the energy of the output data of the intermediate layer to obtain an energy value. A threshold judgment is made on the energy value, and the output of the intermediate layer is selected to be input to the next intermediate layer or the early exit classifier based on the energy value. The early exit classifier performs direct inference analysis to obtain the inference analysis result. The energy prediction model described therein employs a lightweight fully connected layer.

7. The method according to claim 1, characterized in that, In the sequence neural network, a hybrid inference layer is set between the intermediate layers. The hybrid inference layer converts the output data of the intermediate layers into symbolic logic assertions, performs forward chain inference on the symbolic logic assertions to obtain new symbolic conclusions, re-encodes the new symbolic conclusions and symbolic logic assertions, and merges the encoding results with the output data of the intermediate layers as the input data of the next intermediate layer.

8. The method according to claim 1, characterized in that, The sequence neural network is also configured with several branch structures, each including different branches for processing multimodal data. Each branch is a sequence neural network structure. Each branch fuses the output data of the corresponding intermediate layer and outputs it to the next intermediate layer. Between output layers, the output results of different branches are fused.

9. A reasoning analysis acceleration system based on sequence neural networks, characterized in that, Used to perform the method described in any one of claims 1-8.

Images