Research On Neural-symbolic Systems Based On Statistical Relational Learning

Intelligent systems often require both perceptual and reasoning abilities to achieve general artificial intelligence.In recent years,significant progress has been made in perception with neural networks,but their reasoning capabilities remain relatively weak.Conversely,symbolic systems excel at reasoning but exhibit lower learning efficiency.Given the complementary nature of neural networks and symbolic systems,integrating the two has become a key scientific challenge in the research of intelligent systems.Researchers have proposed the approach of neural-symbolic systems to integrate neural networks and symbolic systems.However,existing studies mostly involve either using neural networks to assist symbolic reasoning or symbolic reasoning to assist neural networks,failing to fully leverage the strengths of both.Statistical relational learning,a technique integrating statistical and relational models,is suitable for establishing connections between neural networks and symbolic reasoning.However,the challenge lies in how to utilize this technique to model neural networks and symbolic reasoning as a unified and mutually reinforcing framework.Therefore,this paper starts from the research perspective of using statistical relational learning to integrate neural networks and symbolic reasoning,focusing on neural-symbolic systems for image processing.The main research content and contributions of this paper are as follows:1.A neural-symbolic system framework based on statistical relational learning.Existing neural-symbolic systems face a challenge in balancing the relationship between neural networks and symbolic reasoning.Typically,they either utilize neural networks to assist symbolic reasoning or employ symbolic reasoning to aid neural networks,failing to fully leverage the advantages of both.To address this issue,this paper proposes a Statistical Relational Learningbased Neural-Symbolic System Framework(SRL-NS).SRL-NS aims to unify deep learning models with continuous vector representations and symbolic reasoning with discrete symbol representations into probabilistic computations through statistical relational learning techniques,forming a unified and mutually reinforcing framework.In the SRL-NS framework,symbolic reasoning assists deep learning by making its predictions more logical,aligned with common sense,and easier to interpret,thereby enhancing its generalization ability.Simultaneously,deep learning aids symbolic reasoning to improve its efficiency and robustness to noise.SRL-NS comprises neural reasoning modules and symbolic reasoning modules.The neural reasoning module is a deep learning model responsible for extracting information from unstructured data,while the symbolic reasoning module is a probabilistic graph model responsible for inference.In SRL-NS,the training and testing processes are referred to as concept learning and concept manipulation,respectively.In the concept learning phase,the model learns the basic concepts of symbolic knowledge from unstructured data.In the transductive and inductive concept manipulation phases,the model combines learned basic concepts to generate new concepts,achieving concept generalization in new tasks and scenarios.In experiments involving transductive and inductive manipulation,SRLNS demonstrates outstanding performance,generalization,and interpretability.Importantly,SRL-NS,as a universal framework,exhibits high flexibility.Researchers can design customized neural reasoning modules and symbolic reasoning modules based on the SRL-NS framework to optimize and adapt them to different downstream tasks.In the subsequent research,this paper will use classic tasks in image processing such as visual relationship detection and zero-shot image classification to fully illustrate the advantages of SRL-NS.2.A neural-symbolic system for visual relationship detection.Currently,visual relationship detection methods based on deep learning are facing the problem of insufficient annotated samples,leading to relatively poor predictive performance and a lack of interpretability in the predictions.In order to address this issue,this paper proposes a Bi-level Probabilistic Graph Reasoning method(BPGR)based on the SRL-NS framework.BPGR instantiates the neural reasoning module as a visual relationship detection model and models the symbolic reasoning module as a bi-level probabilistic graph model.The upper layer of the bi-level probabilistic graph model contains the prediction results of the visual relationship detection model based on deep learning,while the lower layer consists of a Markov logic network composed of first-order logic.Through this dual-layer modeling structure,symbolic knowledge can be used to correct errors in the predictions of the visual relationship detection model,thereby significantly improving the predictive performance of the model and achieving interpretability of the prediction results.In the concept learning phase,BPGR learns objects and relationship predicates from unstructured image data.In the concept manipulation phase,BPGR utilizes the predicted results of the well-learned visual relationship detection model on one hand,and explains the predicted results through the bi-level probabilistic graph model on the other hand.Experimental results show that compared to other comparative methods,BPGR has significant advantages in improving predictive performance.More importantly,BPGR enhances the interpretability of the model by providing evidence chains for the predicted results,making the predictions more trustworthy.3.A neural-symbolic system for zero-shot image classification.Current zero-shot image classification methods struggle to effectively capture and explicitly model discriminative features of the target,resulting in weak performance and lack of interpretability when identifying unseen classes.To address this issue,this paper proposes an innovative zero-shot image classification method based on the SRL-NS framework,called Zero-Shot Classification with Logic adapter and Rule prompts(ZSCLR).ZSCLR instantiates the neural reasoning module as a zero-shot image classification model and models the symbolic reasoning module as a logic adapter containing a Markov logic network.ZSCLR is capable of expressing symbolized discriminative features as first-order logic rules and integrates them into the model’s learning process in the form of prompts through the Markov logic network in the logic adapter.This innovative design not only enhances the model’s performance but also endows it with reasoning capabilities,making its decision-making process more transparent and interpretable.In the concept learning phase,ZSCLR learns attributes and class labels of seen classes from unstructured image data.In the concept manipulation phase,ZSCLR utilizes first-order logic rules learned from combined attribute labels of seen classes to infer class labels of unseen classes.Experiments demonstrate that compared to contrastive methods,ZSCLR exhibits significant performance advantages.Furthermore,it also provides interpretability,making the model’s decision-making process more understandable.