This paper introduces a novel approach to program optimization leveraging the power of Graph Neural Networks (GNNs). The core idea is to represent program abstractions, such as Control Flow Graphs (CFGs) and Data Dependency Graphs (DDGs), as graphs and utilize GNNs to automatically identify optimization opportunities and generate optimized code. Traditional program optimization techniques often rely on hand-crafted heuristics and domain-specific knowledge. This approach offers a more automated and potentially more effective method by allowing the GNN to learn complex optimization patterns directly from the program's structure. We demonstrate the feasibility and potential of this approach through theoretical analysis and a conceptual framework. The key contribution lies in the application of GNNs to program optimization, providing a fundamentally new paradigm for achieving better code optimization. ---
Jincheng Zhang· Zenodo (CERN European Organi...· 0 citations
This paper introduces a novel dynamic probabilistic graph model, leveraging high-dimensional neural networks, for natural language processing. We propose a mechanism that dynamically adjusts the network parameters to enhance both semantic reasoning and knowledge graph construction capabilities. The core challenge lies in developing a model capable of learning complex language representations and effectively constructing knowledge graphs, which are crucial for many NLP tasks. The study investigates the effectiveness of this approach through extensive experimentation on various semantic reasoning and knowledge graph tasks. The results demonstrate a significant improvement in these tasks, highlighting the potential of this dynamic probabilistic graph model as a foundation for advanced NLP applications.
Jincheng Zhang· Zenodo (CERN European Organi...· 0 citations
This paper proposes a novel approach to logic programming that integrates neural networks with symbolic reasoning through the dynamic construction and evolution of a knowledge graph. The core idea is to leverage the pattern recognition capabilities of neural networks to learn logical rules and subsequently apply these rules to a knowledge graph. A reinforcement learning agent is employed to guide the dynamic update of the knowledge graph, incorporating new evidence and inferred relationships. This system represents a shift from static knowledge representation to a continuously learning and adapting system. The system's performance is evaluated through simulations, demonstrating its ability to acquire and reason with complex rule sets. The key contribution lies in the synergistic combination of neural and symbolic approaches, offering a robust framework for intelligent reasoning and knowledge acquisition. The system aims to address the limitations of traditional logic programming (e.g., difficulty in handling uncertainty and scaling to complex domains) and the limitations of neural networks (e.g., lack of explainability and difficulty in incorporating prior knowledge). The proposed architecture provides a path towards more flexible and adaptable intelligent systems.
Jincheng Zhang· Zenodo (CERN European Organi...· 0 citations
Graph Neural Networks (GNNs) have emerged as a powerful tool for learning representations from graph-structured data, finding applications in diverse domains like social network analysis, drug discovery, and recommendation systems. However, training GNNs often requires access to sensitive data, raising significant privacy concerns. This work introduces a novel differential privacy mechanism specifically designed for training GNNs. Our approach adds calibrated noise to the gradients during the training process, providing a quantifiable level of privacy protection while maintaining the network's ability to learn effective graph representations. We demonstrate the feasibility and effectiveness of this method through theoretical analysis and conceptual design. The core claim of this work is to develop a differential privacy mechanism that protects the privacy of the training data while still allowing graph neural networks to learn effective representations. The core mechanism involves adding noise to the gradients during the training process, carefully calibrating the noise level to balance privacy and utility. This addresses the privacy challenges of training GNNs.
Jincheng Zhang· Zenodo (CERN European Organi...· 0 citations
Reach audiences
Advertise in front of researchers, engineers, and readers.
Graph embeddings have become a cornerstone of modern graph analysis, enabling the representation of graph structures in continuous vector spaces. However, traditional graph embedding methods often struggle to effectively capture complex dependencies, particularly long-range dependencies and hierarchical relationships inherent in many real-world graphs. This work proposes a novel approach—Contextual Graph Embeddings with Hierarchical Temporal Memory (CGH). We leverage the principles of Hierarchical Temporal Memory (HTM), a biologically inspired neural network architecture designed for learning and representing hierarchical temporal patterns, to enhance graph embeddings. The core idea is to integrate graph embedding techniques with HTM, allowing the model to learn contextual representations that simultaneously capture spatial and temporal relationships within the graph. This results in embeddings that are more robust and informative compared to those generated by conventional methods. We demonstrate the effectiveness of the CGH approach through theoretical analysis and a detailed explanation of the methodology, highlighting its potential for applications in areas such as social network analysis, knowledge graph reasoning, and anomaly detection. The key contribution lies in the fusion of graph embedding's representational power with HTM's ability to model complex, multi-scale patterns.
Jincheng Zhang· Zenodo (CERN European Organi...· 0 citations
This paper introduces a novel neural network architecture, termed "Adaptive Graph Neural Network" (AGN), that leverages dynamic graph structure learning to enhance data modeling and information extraction. Traditional neural networks typically employ static graph structures, limiting their ability to effectively capture complex relationships within data. AGN dynamically adjusts the graph structure during training, adapting to the specific characteristics of the data. This adaptive approach, combined with a novel graph-based representation learning technique, results in improved performance across various data analysis tasks. We present experimental results demonstrating the superior performance of AGN compared to conventional models. The core mechanism focuses on iterative graph refinement guided by a learned loss function that prioritizes information extraction from the graph. This paper explores the potential of dynamically structured neural networks to unlock new capabilities in data analysis and modeling.
Jincheng Zhang· Zenodo (CERN European Organi...· 0 citations
This paper proposes a novel approach to constructing knowledge graphs from implicit, multi-modal data. Traditional knowledge graph construction heavily relies on manual annotation, a process that is both time-consuming and expensive. Our method leverages the rich information contained within user behavior data – encompassing text, images, and videos – to automatically infer hidden knowledge and relationships. We employ deep learning models to represent these diverse modalities within a shared embedding space. Subsequently, we utilize similarity measures or graph neural networks to identify correlations and construct the knowledge graph. This approach significantly reduces the reliance on labeled data and offers a scalable solution for knowledge graph creation, particularly in scenarios where explicit knowledge bases are unavailable or incomplete. The core claim is that analyzing multi-modal user behavior data can effectively build a knowledge graph representing implicit knowledge and relationships. The key mechanism involves mapping different data modalities into a shared vector space using deep learning, followed by relationship discovery through similarity or graph neural networks. This work contributes to a more efficient and automated approach to knowledge graph construction, paving the way for intelligent applications that can reason and understand complex information.
Jincheng Zhang· Zenodo (CERN European Organi...· 0 citations
Existing Relational Neural Networks (R-Nets) demonstrate effectiveness in various graph-based learning tasks; however, their performance often degrades significantly when applied to very large and complex graphs. This paper introduces a novel approach, Hierarchical Graph Decomposition (HGD), integrated within R-Net architectures, to address this scalability limitation. HGD employs a recursive decomposition strategy, breaking down the original graph into progressively smaller subgraphs. This hierarchical structure enables the network to operate on manageable components, improving computational efficiency and maintaining learning performance on large graphs. The core claim is that R-Nets can struggle with very large and complex graphs. The mechanism is the implementation of HGD, which recursively decomposes the graph. This addresses scalability issues of R-Nets via hierarchical graph decomposition.
Jincheng Zhang· Zenodo (CERN European Organi...· 0 citations
This paper introduces a novel approach to Graph Neural Networks (GNNs) termed Contextual Graph Neural Networks with Memory Resonance (CGNMR). The core idea is that GNNs can significantly benefit from incorporating a mechanism that allows nodes to 'remember' and react to previously observed patterns, effectively propagating contextual information across the graph. The proposed CGNMR architecture integrates a recurrent connection within the GNN, creating a dynamic, context-aware representation. This allows nodes to not only process information from their immediate neighbors but also to leverage historical information, leading to improved performance in tasks such as node classification, link prediction, and graph classification. The key contribution lies in the synergistic combination of graph convolutional operations with recurrent memory networks, resulting in a more holistic understanding of graph structures and enhanced predictive capabilities. Mathematical formulations and notation are presented to explicitly define the model and its components.
Jincheng Zhang· Zenodo (CERN European Organi...· 0 citations
This paper explores the integration of high-dimensional data structures and graph neural networks to enhance data analysis and modeling capabilities. The core challenge lies in leveraging the strengths of both paradigms – the inherent connectivity of graphs and the ability of neural networks to learn complex patterns – to achieve significant improvements in efficiency and the discovery of underlying structural and relational insights within high-dimensional datasets. We present a novel framework that combines these elements, utilizing graph neural networks to accelerate analysis and modeling and automatically identify patterns. The paper details the design of the integration strategy, explores the benefits of this fusion approach, and discusses potential applications across various domains.
Jincheng Zhang· Zenodo (CERN European Organi...· 0 citations
This paper presents a novel approach to knowledge graph reasoning utilizing dynamic graph structures. Traditional knowledge graph representation often struggles with the inherent uncertainty and evolving nature of real-world knowledge. We propose a framework that represents knowledge graphs as dynamic graphs, leveraging the power of Graph Neural Networks (GNNs) to perform reasoning tasks. This method allows for the effective capture of relationships and the ability to adapt to changes in the graph structure. The core of our approach lies in modeling entities and relationships as nodes and edges within a dynamic graph, and then employing GNNs to learn the underlying structural patterns and relationships for improved inference. We demonstrate the potential of this dynamic graph representation to enhance both the efficiency and accuracy of knowledge reasoning, addressing key limitations of static knowledge graph approaches. The results indicate a significant improvement in reasoning performance, particularly when dealing with incomplete or evolving knowledge.
Jincheng Zhang· Zenodo (CERN European Organi...· 0 citations
This paper proposes a novel approach to constructing knowledge graphs from implicit, multi-modal data. Traditional knowledge graph construction heavily relies on manual annotation, a process that is both time-consuming and expensive. Our method leverages the rich information contained within user behavior data – encompassing text, images, and videos – to automatically infer hidden knowledge and relationships. We employ deep learning models to represent these diverse modalities within a shared embedding space. Subsequently, we utilize similarity measures or graph neural networks to identify correlations and construct the knowledge graph. This approach significantly reduces the reliance on labeled data and offers a scalable solution for knowledge graph creation, particularly in scenarios where explicit knowledge bases are unavailable or incomplete. The core claim is that analyzing multi-modal user behavior data can effectively build a knowledge graph representing implicit knowledge and relationships. The key mechanism involves mapping different data modalities into a shared vector space using deep learning, followed by relationship discovery through similarity or graph neural networks. This work contributes to a more efficient and automated approach to knowledge graph construction, paving the way for intelligent applications that can reason and understand complex information.
Jincheng Zhang· Zenodo (CERN European Organi...· 0 citations
Assistant Professor Pat Pataranutaporn describes a new interface that lets everyday users glimpse inside an AI's neural network before their chatbot ever says a word.
MIT News · Artificial Intelligence· news.mit.eduJul 6, 2026
PhD student Rachel Sava, winner of the Envisioning the Future of Computing Prize, explores transformative improvements and dystopian risks of neural technology.
