This paper addresses the limitations of current deep learning models in performing abstract symbolic reasoning. The core argument is that deep learning's reliance on pattern recognition hinders its ability to grasp underlying concepts and relationships, leading to a lack of genuine understanding. We propose a novel hierarchical symbolic computation framework designed to overcome these deficiencies. This framework combines the pattern recognition capabilities of deep neural networks with a hierarchical symbolic representation of knowledge. The neural network is tasked with generating symbolic representations from raw input, which are then fed into a symbolic inference engine operating on a hierarchical knowledge graph. This architecture allows for reasoning at multiple levels of abstraction, mimicking the way humans solve complex problems. The key innovation lies in the synergistic integration of these two approaches, fostering a system capable of both learning representations and performing logical deductions. The system's ability to handle ambiguity and context is significantly enhanced through the structured, hierarchical knowledge graph. We outline the system architecture, the training methodology, and the inference process, highlighting the benefits of this hybrid approach. Ultimately, this research contributes to a more robust and flexible approach to artificial intelligence, particularly in domains requiring sophisticated abstract reasoning.
Jincheng Zhang· Zenodo (CERN European Organi...· 0 citations
This paper introduces a novel neuro-symbolic reasoning framework centered around the dynamic construction of a knowledge graph. The core challenge in integrating neural networks and symbolic reasoning lies in the inherent differences in their representations and the difficulty in seamless translation between them. Our approach addresses this by establishing a continuous feedback loop. A neural network generates initial hypotheses, which are then used to construct a knowledge graph. Symbolic rules are then applied to refine and constrain this graph, ensuring consistency and logical validity. This dynamic process allows for a more robust and interpretable knowledge representation, moving beyond the limitations of static knowledge graphs. We detail the architecture, the ruleset, and the interaction mechanisms, emphasizing the system's ability to adapt and learn from both neural and symbolic sources. The system's performance is evaluated through a series of reasoning tasks, demonstrating its effectiveness in complex scenarios.
Jincheng Zhang· Zenodo (CERN European Organi...· 0 citations
This paper proposes a novel graph neural network (GNN) optimization algorithm based on a dynamically adjusted attention mechanism. Traditional GNN optimization methods often rely on fixed architectures, limiting their adaptability to data variations. Our proposed algorithm leverages a self-adaptive strategy to dynamically adjust attention weights, fostering enhanced representation capabilities in GNNs. We introduce a novel attention weighting function that adapts to the data's inherent characteristics, improving the algorithm's robustness and effectiveness in various graph-structured data scenarios. The paper details the algorithm's implementation, provides a comprehensive analysis of its performance through benchmark datasets, and concludes with insights into the potential applications of this approach.
Jincheng Zhang· Zenodo (CERN European Organi...· 0 citations
The goal of this paper is to develop a novel approach to automated generation using self-adaptive graph neural networks (SA-GNNs). Traditional graph neural networks (GNNs) often struggle with adapting to dynamic data, limiting their effectiveness in complex generative tasks. This work introduces a new model, termed SA-GNN, that dynamically adjusts the network's structure and parameters based on input data, enabling the generation of higher-quality outputs. We demonstrate the efficacy of this approach through a series of experiments utilizing various generative datasets, showcasing superior performance compared to existing state-of-the-art methods. This paper presents a complete framework for automated generation, focusing on a robust and adaptable approach to represent and generate data.
Jincheng Zhang· Zenodo (CERN European Organi...· 0 citations
Reach audiences
Advertise in front of researchers, engineers, and readers.
This paper introduces a novel architecture for Graph Neural Networks (GNNs) designed to effectively handle multi-scale spatio-temporal data. The core concept revolves around decomposing time-series data into multiple scales, each represented by a corresponding graph structure. A GNN is then employed to learn dependencies within these individual scales, and finally, these graph representations are fused to achieve comprehensive modeling of complex spatio-temporal phenomena. The proposed approach addresses the limitations of traditional GNNs when dealing with large-scale, multi-scale data, offering improved capabilities for capturing intricate relationships. We demonstrate the effectiveness of this framework through a theoretical analysis, highlighting its advantages over existing methods. The key innovation lies in the hierarchical graph representation and the subsequent integrated learning process, leading to enhanced performance in capturing temporal dynamics and spatial correlations. This work provides a foundational framework for advanced spatio-temporal modeling applications.
Jincheng Zhang· Zenodo (CERN European Organi...· 0 citations
Graph embeddings are a fundamental tool for representing graph data, offering insights into network structure and relationships. Traditional methods often employ fixed weights, limiting the flexibility and scalability of the representation. This paper introduces a novel non-local graph embedding method that dynamically adjusts the weights of each node based on the global structure of the graph, promoting robust and scalable representation. We propose a neural network architecture that learns node weights based on the graph's connectivity and topology, enabling the method to capture complex relationships and adapt to varying graph characteristics. This approach overcomes the limitations of fixed weight methods, offering a significant advancement in graph embedding techniques. This work aims to provide a more adaptable and robust representation of graph data, facilitating a broader range of applications.
Jincheng Zhang· Zenodo (CERN European Organi...· 0 citations
This paper introduces a novel approach to Neural Architecture Search (NAS) that leverages dynamic topology optimization based on dependency relationships. Traditional NAS methods often suffer from inefficiency and limited generalization due to the use of static search spaces and a lack of consideration for intricate layer-level dependencies within neural networks. Our method addresses these limitations by treating a neural network architecture as a graph, where nodes represent layers and edges represent dependencies. A dynamic search algorithm, guided by a dependency scoring function and topological constraints, iteratively modifies this graph's topology to explore a more efficient and effective search space. The core claim is that this dynamic approach, which explicitly models and optimizes topology, significantly improves both the efficiency and generalization capabilities of NAS. The algorithm incorporates a dependency scoring function that evaluates architecture changes based on dependency relationships and performance metrics, while topological constraints prevent drastic and potentially detrimental architectural alterations. The results demonstrate the effectiveness of this dynamic topology-dependent approach compared to conventional static NAS methods.
Jincheng Zhang· Zenodo (CERN European Organi...· 0 citations
This paper presents a novel neural network structure optimization method leveraging adaptive graph neural networks (GNNs). The core challenge in training GNNs lies in the dynamic adjustment of connection weights, which often requires extensive hyperparameter tuning. We introduce a new framework that dynamically adjusts the GNN's connectivity based on data-driven feature analysis. This adaptation mechanism, termed "Adaptive Weight Adjustment," significantly improves the network's expressiveness and generalization performance. We demonstrate the effectiveness of our approach through experiments on several benchmark datasets, showcasing improvements in both accuracy and convergence speed. The proposed method offers a promising alternative to traditional static network architectures and offers a robust solution for optimizing GNNs.
Jincheng Zhang· Zenodo (CERN European Organi...· 0 citations
Graph neural networks (GNNs) have emerged as a powerful tool for various machine learning tasks involving graph-structured data. However, traditional GNN architectures often struggle with noisy and uncertain graph environments, limiting their performance. This paper proposes a novel self-adaptive graph neural network that dynamically adjusts graph structure to mitigate these challenges. We introduce a novel "self-adaptive adjustment" mechanism, allowing the network to dynamically adjust the graph structure in response to noise and uncertainty. This mechanism is implemented using a combination of techniques including adaptive edge weighting and structural pruning. The results demonstrate that our self-adaptive GNN significantly outperforms existing approaches in scenarios characterized by high noise and uncertainty. This work contributes to the development of more robust and accurate GNNs suitable for a wider range of real-world applications.
Jincheng Zhang· Zenodo (CERN European Organi...· 0 citations
This dataset contains the data files used to train the Graph Neural Network (GNN) model. Tetrahedral.pt — data corresponding to the quarter-filled triangular lattice. Trigonal.pt — data corresponding to the half-filled triangular lattice. Neel.pt — data corresponding to the half-filled square lattice. The files are stored in PyTorch .pt format and contain the input data used for training the GNN model.
Relevance. The solvent accessible surface area (SASA) of amino acid residues is a key characteristic for protein structure analysis, but precise methods for calculating it (e.g., FreeSASA) are computationally expensive. Empirical approximations based on the res-idue interaction network (RIN) graph can provide high speed while maintaining ac-ceptable accuracy. Proposed approach. Three empirical functions for estimating relative SASA are pro-posed: approx_sasa, surface_score, and exp_sasa using the degree of the node in RIN as an argument. We present a comparative study of two approaches to graph construction: the classical Cα-graph (threshold 8 Å) and the graph of heavy atoms (Heavy-Atom Graph, HAG, threshold 5.0 Å). The parameters were calibrated on a sample of 509 protein structures (128,794 residues) from various origins using the true relative SASA calculated by the FreeSASA library. Main results. An extended set of 11 RIN topological features was developed and vali-dated, including basic node characteristics, centrality measures (betweenness, eigen-vector, closeness) and hydrophobic subgraph features. Training ensemble models (Random Forest, XGBoost) with these features made it possible to achieve: Random Forest on HAG: MAE = 0.057 ± 0.033, Pearson r = 0.915 ± 0.080 (best result), Random Forest on Cα graph: MAE = 0.066 ± 0.041, Pearson r = 0.890 ± 0.100. Comparison with GNN. We compared our approach with graph neural networks (GCN, GAT, GraphSAGE). GraphSAGE on HAG showed a result close to Random Forest: MAE = 0.0715, Pearson r = 0.8917, indicating the potential applicability of graph neural net-works when using HAG. GCN and GAT performed significantly worse (MAE = 0.14–0.15, Pearson r = 0.51–0.61). Computational efficiency. Empirical formulas are calculated in 0.008 ms per structure (~26,000× faster than FreeSASA), Random Forest in prediction mode is calculated in 36.5 ms (~6× faster than FreeSASA). HAG construction takes 21 times longer than a Cα graph (279.5 ms vs. 13.3 ms). Practical significance. The proposed empirical features are recommended for large-scale pipelines critical to speed and interpretability. Random Forest on HAG is the optimal choice for tasks that require maximum accuracy (MAE = 0.057, Pearson r = 0.915). GraphSAGE on HAG can be considered as an alternative when using deep learning.
Andrey Timofeev, Alexander Bratchikov, Alexander Anufriev· Physchem· 0 citations
Abstract Machine-learned interatomic potentials (MLIPs) have become an increasingly important tool for molecular dynamics (MD) simulations, enabling near quantum-mechanical accuracy at significantly reduced computational cost. Recent studies indicate that the Graph Atomic Cluster Expansion (GRACE) neural network architecture delivers strong performance in materials chemistry. In this work, we assess the GRACE architecture for the prediction of potential energy surfaces for organic molecules and introduce GRACE-OFF (GRACE Organic Force Field). GRACE models of varying depth (one-layer and two-layer) and size (small, medium, large) are trained on the SPICE v2.0 data set. We validate the resulting models using a variety of benchmarks. These include single-point energy and force predictions, torsional energy profiles, condensed phase properties of organic liquids and water (thermodynamic properties, self-diffusion coefficients, radial distribution functions, and temperature-dependent water density), as well as the stability of biomolecular MD simulations for gas-phase Ala15 and solvated crambin. For the single-molecule benchmarks (single point energies and forces, torsional energy profiles), the one-layer models showed only mediocre performance, whereas the two-layer models outperformed the MACE-OFF models to which we compare. For the condensed phase properties, the two-layer models gave consistently better results than the MACE-OFF family of MLIPs. For water and hexane, the GRACE-OFF models also beat the much more expensive small UMA/OMol25 (S) model. The two-layer GRACE-OFF models accurately reproduce experimental water radial distribution functions and predict water densities in close agreement with experimental data over a temperature range from 270 to 330 K. Benchmarks demonstrate that GRACE-OFF achieves higher MD performance than comparable MACE-OFF models in both single and double precision. This establishes GRACE-OFF as an accurate and computationally efficient foundation potential for routine simulations of organic liquids and biomolecular systems.
Anna Katharina Picha, Johannes Karwounopoulos, Linus C. Erhard et al.· Journal of Chemical Theory a...· 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.
