dual-branch architecture design

Existing neural networks are constrained by hierarchical tree-like architectures, which preclude direct communication among sibling nodes and prohibit backward signal propagation to higher-level modules—resulting in weak inter-module collaboration and inefficient representation learning. To address these limitations, we propose the Synchronous Graph Neural Architecture (SGNA), organizing neural units into a modular, dynamically collaborative graph structure that enables arbitrary node-to-node communication and cross-layer signal transmission. Our key contributions are threefold: (1) introducing the first modular graph-structured paradigm for neural architecture design; (2) developing a systematic regularization framework to enforce module independence and load balancing; and (3) generalizing neural architecture search (NAS) to the space of directed acyclic graphs (DAGs). Extensive multi-task experiments demonstrate that SGNA significantly outperforms deep stacked baselines under parameter constraints, achieving superior collaborative representation capability and more comprehensive coverage of the search space.

This work addresses three key limitations of artificial neural networks: the absence of biologically inspired internal neuronal states, selective inter-neuronal communication, and self-organizing topological structure. To this end, we propose Intelligent Neural Networks (INNs), wherein neurons are modeled as first-class entities endowed with memory and online learning capabilities, and rigid layering is replaced by a fully connected graph topology. We introduce two core innovations: (i) a selective state-space model enabling neuron-specific state evolution, and (ii) an attention-guided routing mechanism that facilitates autonomous neuron activation and dynamic, context-aware communication. These design choices significantly improve training stability and model interpretability. On the Text8 benchmark, INNs achieve 1.705 bits per character (BPC), outperforming standard Transformers and matching optimized LSTMs. Notably, a Mamba-based baseline with comparable parameter count fails to converge, empirically validating the critical role of the graph-structured topology in stabilizing training.

Efficiently and accurately approximating complex functions—particularly those exhibiting high oscillation and pronounced local features—remains a fundamental challenge in neural approximation. Method: This paper proposes a Structured-Balanced Multi-Component Multi-Layer Neural Network (MMNN). Adopting a “divide-and-conquer” strategy, MMNN adaptively decomposes the target function into multiple subcomponents, each modeled by a dedicated single-layer subnetwork; hierarchical functional decomposition enables synergistic representation across components. Contribution/Results: MMNN introduces a novel balanced architecture that drastically reduces parameter count (by several-fold compared to fully connected networks or standard MLPs) while preserving strong expressive power. It captures local features precisely without requiring explicit regularization. Experiments demonstrate that MMNN significantly outperforms conventional MLPs in training efficiency, approximation accuracy—especially for highly oscillatory functions—and generalization, effectively breaking the traditional accuracy–efficiency trade-off bottleneck.

Conventional feedforward neural networks are constrained by strictly layered architectures, where neurons within a layer are disconnected, thereby limiting lateral interaction and intra-layer information integration. Method: This paper proposes CHNNet, the first fully connected artificial neural network architecture systematically incorporating intra-hidden-layer lateral connections. It employs intra-layer weight sharing and optimized gradient propagation paths to enhance dynamic inter-neuron interaction and intra-layer integration. Contribution/Results: We provide a theoretical proof that CHNNet’s convergence rate is strictly superior to that of standard feedforward networks. Empirical evaluations across multiple benchmark tasks demonstrate that CHNNet significantly accelerates convergence while improving generalization stability and training efficiency. The core innovation lies in breaking the hierarchical constraint to establish a provably convergent intra-hidden-layer connectivity paradigm—marking a fundamental departure from traditional architectural assumptions in deep learning.

This work addresses the robustness of hierarchical concept representation in spiking neural networks (SNNs) under input occlusion and stochastic neuronal failure. We propose a multi-neuron cooperative coding mechanism that uniformly supports three canonical hierarchical architectures: high- and low-connectivity feedforward networks, and networks with intra-layer lateral connections. Our method integrates multi-representative-neuron encoding, probabilistic recognition modeling, a neuroassembly-inspired lateral learning algorithm, and an extended spiking learning framework. Theoretical analysis and experiments demonstrate that classification accuracy monotonically improves with the number of representative neurons and remains above 92% even under 30% neuronal failure. To our knowledge, this is the first work to simultaneously achieve fault-tolerant hierarchical concept recognition, guaranteed learnability, and quantitatively verified robustness across diverse network topologies—establishing a novel paradigm for reliability-aware design of brain-inspired perception systems.

This work addresses the diminished understanding of neural network fundamentals caused by the widespread use of high-level deep learning libraries. To bridge this gap, the authors construct a complete neural network framework from scratch, eschewing automatic differentiation and prebuilt modules. The implementation explicitly details forward and backward propagation, incorporates multiple activation functions, L2 regularization, and advanced optimizers such as Adam. Designed to balance pedagogical clarity with engineering scalability, the framework demonstrates numerical stability, correctness, and generalization capability on multiclass classification tasks. It thus provides a reproducible and extensible tool for both research and instruction, fostering deeper insight into the core principles of deep learning.

This study addresses the lack of general design principles linking neural network architecture to computational capacity. By systematically evaluating the computational performance of recurrent neural networks with diverse connectivity patterns on Boolean function tasks through large-scale sampling, the work reveals— for the first time—that local 2-cycles and 3-cycles are critical structural motifs for enhancing computational power. It further demonstrates that introducing a small number of sparse connections and biologically inspired interneuron-like units significantly boosts the performance of large-scale networks. The authors construct a comprehensive performance map linking small-network architectures to Boolean function realization, showing that networks containing short cycles achieve optimal performance. Moreover, network performance can be accurately predicted from structural statistics, offering a theoretical foundation and biologically inspired guidance for future neural architecture design.

This work addresses the efficient deployment of spiking neural networks (SNNs) on neuromorphic hardware by optimizing the mapping of neurons to hardware cores to minimize communication overhead and resource consumption. For the first time, SNNs are modeled as hypergraphs, where hyperedges explicitly capture spike duplication among neurons within a core, revealing that hyperedge overlap and locality are critical for high-quality mappings—an insight beyond the capability of conventional graph models. Building on this formulation, we design and enhance hypergraph partitioning and placement algorithms to enable efficient mapping of large-scale, biologically plausible SNNs under diverse execution time constraints. Experimental results demonstrate that our approach significantly outperforms state-of-the-art methods, substantially reducing both communication traffic and hardware resource usage while supporting scalable deployment of SNNs of arbitrary size.

This work addresses the inefficiencies of conventional neural networks in data and energy consumption, as well as the limited availability of effective learning algorithms for spiking neural networks (SNNs). To bridge this gap, the authors propose Spark, a modular SNN framework that constructs end-to-end models by composing simple plasticity-based components, enabling continuous, batch-free learning. Spark integrates seamlessly with traditional machine learning pipelines while incorporating biologically inspired continual learning mechanisms, substantially improving data efficiency and practical applicability. The framework’s effectiveness is demonstrated on the sparse-reward CartPole task, where it successfully learns in a continuous setting, highlighting the promise of modular design for efficient SNN training.