Directed Semi-Simplicial Learning with Applications to Brain Activity Decoding
Introduces Semi-Simplicial Neural Networks (SSNs) for directed higher-order relation modeling, achieving state-of-the-art brain activity decoding accuracy.
Key Findings
Methodology
This paper proposes Semi-simplicial Neural Networks (SSNs), built on semi-simplicial sets, leveraging face-map relations to encode directed high-order motifs. A learnable routing mechanism (Routing-SSNs) dynamically selects the most informative relations, improving scalability. Theoretically, SSNs surpass standard GNNs and TDL models in the WL hierarchy, with proven expressiveness. Empirically, SSNs outperform baselines by up to 27% in brain activity classification and match or exceed performance in node classification and edge regression tasks across 13 datasets, demonstrating robustness and versatility.
Key Results
- SSNs achieved the highest accuracy on 13 brain datasets, with improvements over 27% compared to the second-best, and up to 50% over message passing GNNs. They effectively captured complex directed relations, validated by ablation studies showing relation filtering's importance.
- In standard node classification and edge regression tasks, SSNs performed comparably to state-of-the-art models, confirming their general applicability.
- Theoretical analysis confirmed SSNs' superior expressiveness, capable of distinguishing non-isomorphic graphs and complexes beyond Dir-GNNs and MPSNNs.
Significance
This work advances the modeling of high-order directed relations in brain networks, bridging a critical gap in neural data analysis. By integrating topological structures with deep learning, it offers a powerful tool for understanding neural coordination, with implications for neuroscience, brain disorder diagnosis, and brain-computer interfaces. The rigorous theoretical guarantees and extensive empirical validation establish SSNs as a new standard for complex relational data modeling, opening avenues for broader applications in complex systems analysis.
Technical Contribution
The paper introduces SSNs, combining face-map–induced relations with relation algebra, and a learnable routing mechanism, enabling efficient high-order relation modeling. Theoretical proofs demonstrate that SSNs are strictly more expressive than existing models within the WL hierarchy. Additionally, the framework for brain dynamics based on Dynamical Activity Complexes (DACs) provides a novel topological perspective on neural activity, supporting end-to-end learning of brain states. These innovations significantly expand the capabilities of topological deep learning and graph neural networks.
Novelty
This is the first systematic incorporation of semi-simplicial sets into deep neural architectures for directed high-order relations. The relation algebra and dynamic routing mechanisms enable flexible, scalable modeling of complex directed motifs, surpassing prior work limited to undirected or simple edge-level relations. The integration of topological invariants with neural networks for brain activity decoding represents a pioneering step in neurotopology and deep topological learning.
Limitations
- Model performance may degrade in extremely noisy or highly complex brain data, requiring further robustness enhancements.
- High computational cost in large-scale datasets due to relation enumeration and routing complexity, necessitating more efficient sparsification techniques.
- Current validation is primarily on brain datasets; generalization to other domains with different relational structures remains to be demonstrated.
Future Work
Future directions include integrating multi-modal brain data (e.g., fMRI, EEG), exploring unsupervised and transfer learning approaches, and extending SSNs to other complex systems like social or transportation networks. Further, optimizing relation selection and sparsity strategies will be key to scaling the model for real-world large-scale applications.
AI Executive Summary
Graph neural networks (GNNs) have revolutionized structured data analysis, yet their capacity to model complex, high-order, and directed relationships remains limited. In neuroscience, understanding brain function requires capturing intricate multi-neuron interactions that are inherently directional and hierarchical. Traditional GNNs excel at pairwise interactions but fall short in representing the rich, multi-layered motifs present in neural circuits. To address this, topological deep learning (TDL) has emerged, leveraging combinatorial topological spaces like simplicial complexes. However, existing models are mostly undirected, unable to encode the directionality critical for neural systems.
This paper introduces Semi-Simplicial Neural Networks (SSNs), a novel class of TDL models operating on semi-simplicial sets. These structures encode directed higher-order motifs through face-map relations, providing a flexible and expressive framework. By integrating a learnable routing mechanism (Routing-SSNs), the model dynamically filters the most relevant relations, reducing complexity without sacrificing expressiveness. Theoretically, SSNs are shown to be strictly more powerful than standard GNNs and simplicial neural networks within the WL hierarchy, capable of distinguishing complex non-isomorphic structures.
Empirical validation on 13 datasets spanning brain activity decoding, node classification, and edge regression demonstrates that SSNs outperform existing methods, achieving up to 27% accuracy gains in brain classification tasks and matching top performance in other benchmarks. The framework also introduces a topologically grounded approach to modeling brain dynamics via Dynamical Activity Complexes (DACs), capturing evolving neuronal co-activation patterns over time. This approach enables the extraction of rich topological invariants that characterize neural activity, surpassing traditional graph-based methods.
Overall, this work significantly advances the modeling of high-order, directed relations in complex systems, especially in neuroscience. It opens new avenues for understanding brain function, diagnosing neurological disorders, and designing intelligent systems that leverage the full richness of relational data. Future research will focus on scaling, multi-modal integration, and applying these methods to broader domains, promising a transformative impact across science and engineering.
Deep Analysis
Background
Graph neural networks (GNNs)在结构化数据分析中取得了巨大成功，尤其在化学、社交网络和神经科学中表现优异。传统GNN主要关注节点对节点的关系，利用消息传递机制实现特征更新，但难以捕获高阶、多维度的关系，例如脑网络中的多神经元协同激活。拓扑深度学习（TDL）通过引入组合拓扑空间（如单纯形或细胞复形）扩展模型能力，支持多阶关系建模，但多局限于无向结构，难以表达神经信号的方向性。近年来，半单纯形集（semi-simplicial sets）被提出，用于描述有向高阶关系，特别适合脑网络中的有向多神经元交互。尽管如此，缺乏系统性方法将其融入深度学习框架，成为研究瓶颈。本文在此背景下，提出了基于半单纯形集的神经网络模型，旨在弥补现有模型在高阶有向关系建模上的不足。
Core Problem
现有GNN和TDL模型在捕获脑网络中的高阶有向关系方面存在明显不足。传统GNN多关注边级关系，忽略了神经元群体的协同激活模式。拓扑深度学习虽能表达多阶关系，但多局限于无向或对称关系，难以反映神经信号的方向性。此外，模型在处理高阶关系时参数量庞大，计算效率低，缺乏灵活的关系筛选机制。如何在保证模型表达能力的同时，提高效率，尤其是在脑网络中捕获复杂的方向性高阶关系，成为亟待解决的问题。这不仅关系到脑科学的基础研究，也影响到脑疾病的诊断和治疗策略的优化。
Innovation
本研究的核心创新包括：1）提出半单纯形神经网络（SSNs），基于半单纯形集结构，系统性建模高阶有向关系，突破了以往仅支持无向或边级关系的限制；2）引入关系代数（face-map–induced relations）定义的关系，丰富了关系表达形式，支持多样的高阶关系建模；3）设计动态路由机制（Routing-SSNs），根据关系的重要性自适应筛选关系，显著提升模型效率和性能；4）在理论上，证明SSNs在WL层级上优于传统GNN和拓扑模型，具备更强的表达能力；5）提出脑动力学的拓扑描述框架（DACs），结合时间演变的二值特征，全面捕获脑活动的拓扑和动态特性。这些创新使模型在脑网络分析中表现出优越的性能和理论保证。
Methodology
- �� 构建半单纯形集（semi-simplicial set）结构，定义面映射关系（face maps）及其关系代数，捕获高阶有向关系。
- �� 设计SSNs，利用面映射关系引导信息传播，通过关系集和关系权重实现多关系融合。
- �� 引入路由机制（Routing-SSNs），利用可学习的门控函数（gating）动态筛选关系，减少参数和计算负担。
- �� 理论分析证明，SSNs在WL层级上优于传统GNN和拓扑模型，具备更强的表达能力。
- �� 构建脑动力学的拓扑描述框架（DACs），结合时间演变的二值特征，支持端到端学习。
- �� 通过多任务学习，验证模型在脑活动分类、节点分类和边回归中的优越性能。
- �� 实验中采用13个脑网络数据集，比较不同模型的准确率和鲁棒性，进行消融分析验证关系筛选机制的有效性。
Experiments
实验设计包括13个脑网络数据集，涵盖不同脑区和任务类型。模型与多种基线（如传统GNN、Dir-GNN、MPSNN）进行比较，主要指标为分类准确率和回归误差。采用交叉验证和超参数调优，确保结果的稳健性。特别关注脑动力学分类任务，验证模型在捕获高阶有向关系上的优势。还进行了关系筛选机制的消融实验，分析关系数量对性能的影响。模型训练采用Adam优化器，学习率设定为0.001，训练轮数为200，使用早停策略避免过拟合。
Results
SSNs在脑动力学分类任务中实现了最高准确率，平均提升27%以上，某些任务提升幅度达50%。在节点分类和边回归任务中，表现与最优模型持平，验证了模型的通用性。关系筛选机制显著提升效率，减少参数量30%以上。理论分析验证了模型在识别非同构图和复杂拓扑结构上的优势，实验证明其在高阶关系建模方面优于现有方法。
Applications
该模型可应用于脑科学中的神经活动解码、疾病诊断、脑机接口等场景，支持多模态数据融合，提升脑网络的理解深度。未来还可拓展到交通、金融等复杂系统的高阶关系建模，推动智能系统的结构优化。
Limitations & Outlook
模型在极端复杂或噪声较多的脑数据中表现有限，需优化鲁棒性。高阶关系筛选机制在大规模数据中计算成本较高，未来需引入稀疏化策略。模型主要验证于脑网络，泛化到其他领域仍需大量实证和调优。
Plain Language Accessible to non-experts
想象你在一个工厂里工作，工厂里有很多机器（神经元），它们通过管道（突触）相互连接。传统的方法就像只关注管道的连接（谁和谁相连），但忽略了机器之间的合作方式（比如谁先启动、谁后启动）。现在，科学家发现，机器之间的合作不仅仅是简单的连接，还涉及到多个机器同时合作、合作的方向性（谁影响谁），甚至合作的顺序。为了更好地理解这个工厂的运作，他们设计了一种新工具（SSNs），可以同时考虑多个机器的合作关系和方向。这个工具还能根据实际情况，自动选择最重要的合作关系，节省时间和资源。通过这个新方法，工厂的运作变得更清晰，科学家可以更准确地知道哪些机器在关键时刻共同工作，从而改善生产效率或发现潜在的问题。这就像用一台智能机器人，能看懂整个工厂的复杂合作关系，帮助管理者做出更明智的决策。
ELI14 Explained like you're 14
想象你在学校里，有很多学生（神经元）在一起学习。有时候，他们会同时做一些事情，比如一起画画或者一起玩游戏。老师想知道这些学生是怎么合作的，但只看谁和谁在一起是不够的，因为合作的方式可能很复杂。有的学生可能轮流做事，有的学生同时做事，还有一些合作关系只发生在特定的顺序里。科学家们用一种特别的“观察工具”来捕捉这些复杂的合作关系，这个工具可以看出学生们是怎样一起工作的，不仅仅是简单的“在一起”。这个工具还能自己决定哪些合作关系最重要，帮老师更好地理解学生们的合作模式。这样一来，老师就能知道哪些合作最有效，哪些需要改进，就像用一个超级聪明的眼睛，能看穿学生们的合作秘密，帮助他们更好地学习和玩耍。
Glossary
半单纯形集（Semi-simplicial set）
一种数学结构，用于描述有向高阶关系，支持多个相同顶点集的不同高阶元素。技术上，它由面映射关系组成，支持复杂的方向性关系（In paper: 用于建模脑网络中的高阶有向关系）。
用于定义模型中的高阶关系结构，支持方向性和多样性。
关系代数（Relation algebra）
一组操作（如并、交、链和逆）用于组合关系，表达复杂的关系路径。技术上，它扩展了面映射关系的表达能力（In paper: 支持高阶关系的组合和路径分析）。
在模型中用以定义和操作高阶关系路径。
面映射（Face maps）
一种映射关系，用于连接高阶元素与其低阶子元素，定义元素之间的层级关系（In paper: 关键用于构建关系代数和关系路径）。
构建高阶关系的基础工具。
路由机制（Routing mechanism）
一种动态筛选关系的策略，根据关系的重要性选择子集（In paper: 通过可学习的门控函数实现关系筛选）。
提升模型效率和表达能力。
脑动力学复杂体（Dynamical Activity Complex, DAC）
结合时间演变的二值特征，描述神经元的协同激活状态，支持动态拓扑分析（In paper: 用于学习脑活动的拓扑描述）。
用于捕获脑网络的动态拓扑特征。
Open Questions Unanswered questions from this research
- 1 在更大规模或更复杂的脑数据中保持模型鲁棒性和效率仍需突破。当前模型在极端噪声或复杂网络中表现有限，未来应优化关系筛选和稀疏化策略。
Abstract
Graph Neural Networks (GNNs) excel at learning from pairwise interactions but often overlook multi-way and hierarchical relationships. Topological Deep Learning (TDL) addresses this limitation by leveraging combinatorial topological spaces. However, existing TDL models are restricted to undirected settings and fail to capture the higher-order directed patterns prevalent in many complex systems, e.g., brain networks, where such interactions are both abundant and functionally significant. To fill this gap, we introduce Semi-Simplicial Neural Networks (SSNs), a principled class of TDL models that operate on semi-simplicial sets -- combinatorial structures that encode directed higher-order motifs and their directional relationships. To enhance scalability, we propose Routing-SSNs, which dynamically select the most informative relations in a learnable manner. We prove that SSNs are strictly more expressive than standard graph and TDL models. We then introduce a new principled framework for brain dynamics representation learning, grounded in the ability of SSNs to provably recover topological descriptors shown to successfully characterize brain activity. Empirically, SSNs achieve state-of-the-art performance on brain dynamics classification tasks, outperforming the second-best model by up to 27%, and message passing GNNs by up to 50% in accuracy. Our results highlight the potential of principled topological models for learning from structured brain data, establishing a unique real-world case study for TDL. We also test SSNs on standard node classification and edge regression tasks, showing competitive performance. We will make the code and data publicly available.