Neural Chains: Sequential Architectures

• Neural Chains are sequential feed-forward architectures that encode and propagate information using structured neural modules in biological, artificial, and hybrid networks.
• They integrate dual-chain designs, such as synfire chains and transformer-style networks, to achieve temporally precise and robust signal transmission through graded propagation.
• Applications span from network pruning and classifier design to modeling wave propagation and solving PDEs, underscoring their computational universality and dynamic robustness.

Neural chains refer to a broad class of sequential, feed-forward architectures and dynamic processes—biological, artificial, and hybrid—that encode, propagate, process, or match information in structured series of neural modules. This notion encompasses synfire chains observed in cortex, modular architectures in deep learning, chain-like cores in functional connectomics, and specialized algorithmic chains in network pruning and classifier design. Neural chains are central to understanding how sequential information, synchronous rhythms, and graded signals can be transmitted, processed, and robustly manipulated across layers in both artificial and biological networks.

1. Fundamental Architectures and Mathematical Formulations

Neural chains are typically characterized by sequences of neural modules or populations organized in a feed-forward topology. In the classical synfire chain, each layer consists of a population of neurons whose synchronous collective spiking drives the next layer, enabling temporally precise transmission of activity (Wang et al., 2015). Advanced constructions such as synfire-gated synfire chains (SGSCs) employ two coupled chains—a gating chain producing stereotyped high-amplitude pulses, and a graded chain propagating continuous firing rate amplitudes. The SGSC architecture involves:

• $M$ layers per chain.
• Population sizes $N_1$ (graded chain) and $N_2$ (gating chain).
• Layer-to-layer connectivity matrices $K^{\sigma\sigma'}_{jk}$, connection probabilities $p_{\sigma\sigma'}$, and synaptic strengths $S^{\sigma\sigma'}$.
• Cross-chain gating (gating $\rightarrow$ graded) and logic (graded $\rightarrow$ gating).

The neuronal dynamics typically involve current-based integrate-and-fire units:

$$\frac{d}{dt}\,v^\sigma_{i,j}(t) = -g_{\mathrm{leak}}\big(v^\sigma_{i,j}-V_{\mathrm{leak}}\big) + \sum_{\sigma'} I^{\sigma\sigma'}_{i,j}(t) + I^{\sigma}_{\mathrm{bg}}(t)$$

where $g_{\mathrm{leak}}$ and $V_{\mathrm{leak}}$ are biophysical parameters, with Poisson background input and first-order kinetics for synaptic currents.

In deep learning, a neural chain is an $L$-layer transformer-style feed-forward network (without attention), each layer being a fixed-width vector transform:

$$z^{(k+1)} = f\big(W_k z^{(k)} + b_k\big), \quad k=0,\dots, L$$

where $f$ is a nonlinearity, and $W_k$ are $N \times N$ matrices (Succi et al., 1 Jan 2026).

In classifier design, a neural chain may denote a chain of local binary classifiers correcting the output of a global model, as in Local Classifier Chains CNN (LCC-CNN) (Zhang et al., 2018). In this context, each sample propagates through a series of local binary nets, each disambiguating a pair of labels (selected via similarity or confusion matrices).

2. Information Processing and Dynamical Principles

Neural chains support multiple modes of information transfer:

• Pulse-gated propagation: Synchronous gating pulses from one chain open integration windows in a graded chain, enabling continuous or discrete amplitude transfer (Wang et al., 2015).
• Mean-field graded propagation: Overlapping pulses allow a graded waveform to cascade rapidly across many layers, with translational invariance ensured by precise determinantal conditions on synaptic gain $S$.
• Persistent and transient sequential activity: In Quadratic Integrate-and-Fire networks with temporally asymmetric Hebbian connectivity, persistent synfire-chain dynamics arise as a rotating-wave limit cycle, while transient hippocampal replay emerges via slow modulatory drive and spiral excursions around fixed points (Shimizu et al., 8 Aug 2025).

In discrete-time heterogeneous chain networks, bidirectional coupling and node-type heterogeneity produce rich dynamical phenomena: multistability, period-doubling, Neimark-Sacker bifurcations, and attractor coexistence (periodic and chaotic regimes). Quantitative measures (cross-correlation coefficients, Kuramoto order parameters, sample entropy) reveal partial synchrony and information-processing complexity across parameter regimes (Ghosh et al., 2024).

Traveling-wave generation in unidirectional oscillator chains hinges on precise tuning of local interactions and is mathematically proved to be robust and globally stable under parameter perturbations, supporting families of traveling waves with arbitrary period and wavenumber (Fernandez et al., 2014).

3. Robustness, Modularity, and Pruning in Chain Architectures

Neural chains exhibit robustness to noise and heterogeneity, both in theory and simulation. SGSC models are tolerant to finite-size fluctuations (signal-to-noise $\gg$ 10 after multiple layers), parameter variability (synaptic weights, pulse timing), and maintain high fidelity of amplitude transfer when pulse overlap exceeds $\eta \gtrsim 2$ (Wang et al., 2015).

Synfire chains formed by spike-timing dependent plasticity (STDP) and potentiation decay organize stochastically: fast potentiation decay yields long, variable chains; slow decay produces short, sharply peaked distributions; axon-remodeling ensures group-size stability (Miller et al., 2013). Successive recruitment and chain closure are governed by feedback inhibition and lottery-growth statistics.

In deep learning, graph-theoretic pruning (LEAN) treats a CNN as a DAG of operators, extracting longest chains via dynamic programming on operator norms. This chain-based pruning preserves path coherence, avoids disjointness, enables dramatic reduction in filter count (1.7–12$\times$), and outperforms layer-wise or magnitude-based approaches with minimal impact on accuracy (Schoonhoven et al., 2020).

Classifier chains in LCC-CNN avoid error propagation by chaining local binary nets only when confidence is boosted; multi-link chains are rare, limiting computational cost, and local models focus computation where confusion is highest (Zhang et al., 2018).

4. Probabilistic Graph Models and Theoretical Interpretations

Mapping neural networks to chain graphs (CGs) in probabilistic graphical models provides rigorous semantics for every component (layer, connection type, activation, dropout). A CG is a mixed graph partitioned into chain components with undirected intra-component edges and directed acyclic inter-component edges (Shen et al., 2020).

Feed-forward computation arises as a block-coordinate mean-field inference on the CG, with ordinary layerwise propagation corresponding to single-pass variational free-energy minimization. Partially collapsed feed-forward (PCFF) generalizes dropout, allows explicit sampling from marginal distributions, and achieves competitive regularization (Shen et al., 2020).

Novel architectural insights include:

• Residual and skip connections as refinement modules in the CG.
• Convolutional layers as sparse bipartite CRFs.
• RNNs as time-unrolled chain graphs.
• Activation function choices dictated by accuracy of mean-field approximations.

5. Biological, Functional, and Network-Theoretic Contexts

Functional connectomics analysis of human resting-state networks reveals chain-like cores comprising sequentially interconnected, anatomically homogeneous modules—occipital, cerebellar, parietal, frontal—stitched together via maximum spanning forest and tree extraction (Mastrandrea et al., 2017). This chain-backbone topology economizes wiring length, establishes bridging hubs, and supports graded integration/segregation as coupling thresholds vary.

Central pattern generators (CPGs) and their feedforward lifts formalize rhythmic propagation in animal locomotion and continuum robotics. Stability criteria for synchronous and phase-synchronous lifts are reduced to Floquet multipliers on the CPG core; propagation remains stable in extended chains if all transverse multipliers lie inside the unit circle across standard neuron models (Hodgkin-Huxley, FitzHugh-Nagumo, Morris-Lecar, Hindmarsh-Rose) (Stewart et al., 13 Jun 2025).

6. Computational Universality and Extreme Depth

Chains of width-one perceptrons (deepest neural networks) are formal universal classifiers: any finite decision boundary can be simulated by sufficient depth, with each layer adding a bit of information (Rojas, 2017). This extreme serialization is theoretically universal but practically inefficient—depth scales exponentially with the required boundary complexity, and construction is not gradient-based but combinatorial.

Key implications include:

• Duality to width-based universality results.
• Emphasis on skip connections as forerunners of residual learning.
• Exponential depth–complexity trade-off and lack of feature abstraction in narrow neural chains.

7. Neural Chains in Machine Learning for Dynamics and PDEs

Neural chains, as transformer-style feed-forward networks, are interpretable as discretizations of neural integral and PDE models: each layer represents Euler stepping in time and spatial convolutions approximate local derivatives (Succi et al., 1 Jan 2026). Standard finite-difference (FD) solvers possess unique banded weight matrices; physics-informed neural networks (PINNs), optimized by SGD, typically converge to random dense matrices due to the vast entropy of the solution space, sacrificing explainability and training efficiency.

Compared on 1D Burgers and Eikonal equations, FD chains are minimal and transparent; PINN chains yield similar solutions with $\mathcal{O}(L N^2)$ random parameters, robust but opaque, suggesting possible advantages in high-dimensional regimes where grid-based FD methods suffer from the curse of dimensionality.

---

In summary, neural chains constitute a unifying organizational principle across neuroscience, machine learning, network theory, and applied mathematics, capturing the dynamics, modularity, robustness, and compositionality of sequential information processing. Their manifestations range from biophysically realistic spiking circuits and stochastic plasticity-driven recruitment to feed-forward network pruning, classifier hierarchies, deep learning architectures, and mathematical models of wave propagation and critical connectivity. The neural chain paradigm offers mechanistic explanations, robust computational strategies, and theoretical connections across domains.