Statistical-Neural Interaction (SNI)

• Statistical-Neural Interaction is a framework that distinguishes intrinsic neural connectivity from confounding statistical dependencies using a multi-step inference procedure.
• It employs methods like GLMs, persistent homology, and autoregressive density models to decouple stimulus effects from noise, enhancing model stability and interpretability.
• Applications of SNI span connectomics, causal inference, and advanced neural network training, offering quantifiable improvements in network reconstruction and predictive accuracy.

Statistical-Neural Interaction (SNI) refers to a class of theoretical frameworks and algorithms that rigorously characterize, identify, or leverage the mutual dependencies between statistical structure (including data-driven correlations and noise) and neural interactions (including synaptic couplings, population dynamics, or networked computation) in neuroscientific and machine learning contexts. SNI serves both as an analytic paradigm—for distinguishing extrinsic statistical effects from genuine neural connectivity—and as a principle for model construction, parameter estimation, inference, and phenomenological interpretation spanning neural data analysis, dynamical systems, and neural network theory.

1. Formalism and Identifiability of Neural Interactions

Statistical-Neural Interaction is grounded in the distinction between dependencies that arise from shared input, stimulus design, or global statistical structure and those reflecting actual network couplings or causal influences.

A canonical instantiation is the two-step inference procedure for population generalized linear models (GLMs) of spiking neural populations. The SNI procedure decouples stimulus-driven statistics from intrinsic noise correlations by fitting, in succession:

1. Stimulus filter inference (uncoupled LNP model): For each neuron $i$, fit the stimulus response filter $K_i$ in an uncoupled Poisson (LNP) model, using unrepeated stimuli. The penalized objective is:

$$\mathcal{J}_{\rm stim} = -\sum_{i,t}[n_i(t)\ln\lambda_i^{\rm LNP}(t)-\lambda_i^{\rm LNP}(t)] + \alpha\sum_{i,\tau,x,y}|K_{i,xy}(\tau)| + \beta\sum_{i,\tau,x,y}\|\nabla_{x,y}K_{i,xy}(\tau)\|^2$$

with $$\lambda_i^{\rm LNP}(t) = \exp\{h^i_{\rm offset}+h^i_{\rm stim}(t)\}$$.

2. Coupling inference (“PSTH-clamped” model): With repeated-trial data, hold the peri-stimulus time histogram ($u_i(t)$) clamped to empirical means, and fit only the coupling filters $J_{ij}$ from fluctuations around the mean, using:

$$\mathcal{L}_{\rm coup} = \sum_{i,t}[n_i(t)\ln\lambda_i^{\rm coup}(t) - \lambda_i^{\rm coup}(t)] - \gamma\sum_{i,j,\tau}|J_{ij}(\tau)|$$

with $$\lambda_i^{\rm coup}(t)=\exp\{h_{\rm offset}^i+u_i(t)+h^i_{\rm int}(t)\}$$, ensuring only noise-induced covariances remain.

3. Merged Model: The final GLM expresses:

$$\lambda_i(t) = \exp\Bigl\{h_{\rm offset}^i + h_{\rm stim}^i(t) + \sum_{j,\tau}J_{ij}(\tau)[n_j(t-\tau)-\lambda_j^{\rm LNP}(t-\tau)]\Bigr\}$$

Subtracting $\lambda_j^{\rm LNP}$ aligns the mean, ensuring only intrinsic noise correlations are ascribed to $J_{ij}$ (Mahuas et al., 2020).

This procedure eliminates the confounding between stimulus-induced covariances and genuine synaptic or network couplings, yielding coupling estimates that generalize across stimulus statistics and are robust against simulation blow-ups (self-excitation transients). In practice, it achieves CoD (coefficient of determination) $\approx 0.91$ on moving bar stimuli, compared to $\approx 0.55$ for single-step fitting, and prevents instability under cross-stimulus generalization (Mahuas et al., 2020).

2. SNI in Neuroscientific Data: Stochastic Processes and Network Graphs

Statistical-Neural Interaction is also formulated for stochastic neural network models with memory. Consider a system of interacting chains with history dependence: the spiking of neuron $i$ depends on presynaptic activity since its own last spike, embodying renewal properties. The directed interaction graph $G=(V,E)$ is defined via nonzero synaptic weights $W_{j\to i}$.

To infer $G$, SNI-based estimators compare conditional firing probabilities of neuron $i$ given different histories for candidate presynaptic neuron $j$. The key selection rule is:

$$\Delta_{(i,n)}(j) = \max_{w\in T_{(i,n)}} \max_{v\in T_{(i,n)}^{w,j}} |\hat p_{(i,n)}(1|w)-\hat p_{(i,n)}(1|v)|$$

where $T_{(i,n)}^{w,j}$ is the set of local pasts differing from $w$ only at $j$. Vertices $j$ with $\Delta_{(i,n)}(j)>\epsilon$ are included in the estimated neighborhood. Non-asymptotic exponential bounds for under- and overestimation are provided, and strong consistency is achieved without stationarity assumptions (Duarte et al., 2016). The SNI perspective here rigorously frames interaction graph recovery as identifying statistically influential presynaptic neurons based on empirical sensitivities in transition probabilities.

3. Statistical–Neural Interaction in Connectomics and Causal Inference

SNI underpins contemporary frameworks for functional and causal connectomics. Three key graph types are distinguished:

• Anatomical connectome: Synaptically annotated graph, independent of dynamics.
• Functional connectome (AFC): Encodes statistical associations (covariance, partial correlation, mutual information) among time series $X_v(t)$.
• Causal functional connectome (CFC): Directed graph representing putative “causal” (intervention-predictive) relationships, expressible via Directed Markov Property (DMP) in a DAG:

$Y_v \indep Y_{nd(v)\setminus pa(v)} \mid Y_{pa(v)}$

Association-based SNI approaches—correlation, partial correlation, Markov random fields—recover undirected dependency networks; causal inference methods (Granger causality, DCM, constraint-based graphical models) enforce directionality and intervention semantics (Biswas et al., 2021).

Key inference strategies:

Method	Causality Type	Temporal	Parametricity	Limitations
Granger Causality	Predictive (VAR model)	Yes	Linear	No counterfactual semantics
DCM	Mechanistic	Yes	Nonlinear	Priors, cannot discover structure
Probabilistic Graphical Models	Interventional (do-calculus)	No (default)	Nonparametric	Faithfulness/acyclicity required

SNI thus systematizes the translation from statistical association to causal/mechanistic connectivity, unifying model-based and purely data-driven approaches within a Markov-graphical, conditional-independence framework.

4. SNI in Theoretical and Statistical Physics of Neural Systems

The SNI paradigm is explicit in the statistical mechanics of neural populations. In neural statistical mechanics, one describes network activity by Hamiltonians or energy functions (e.g., Ising, Hopfield) with couplings $J_{ij}$ (synapses) and studies their macroscopic order parameters (e.g., mean overlaps $q$, storage capacity, phase diagrams). Statistical field theory formulations introduce activity fields $\phi(x,t)$ and connectivity fields $\psi(x,y)$, yielding coupled variational equations and propagators encoding SNI:

• Action functional: $S[\phi,\psi]$ incorporates neural dynamics, plasticity, and neural–connectivity coupling via terms such as $-\lambda\int \phi^*(x)\phi(y)\psi(x,y) dxdy$.
• Response (Green’s) functions: The inverse quadratic form $L$ in fluctuations yields propagators $G_{\phi\phi}$ (activity), $G_{\psi\psi}$ (connectivity), and cross-susceptibility $G_{\phi\psi}$ (the SNI kernel).

Explicitly, off-diagonal blocks proportional to $\lambda\phi_0$ in $L$ encode the bidirectional interaction between statistical connectivity fluctuations and neural field fluctuations, quantifying SNI at the mesoscopic or field level (Gosselin et al., 2023, Gosselin et al., 2023).

SNI also structures the statistical-physics analysis of computation in high-dimensional circuits. The replica and cavity methods, random matrix theory, and belief propagation analyze self-averaging laws for order parameters under disordered coupling. Such analyses provide the fundamental ground for algorithms and limits in low-rank inference, compressed sensing, and random-projection dimensionality reduction (Advani et al., 2013).

5. SNI Algorithms in Machine Learning and Neural Networks

Statistical–Neural Interaction is formalized in several modern algorithms for interpreting and utilizing neural networks:

• Neural interaction detection (NID): Statistical interactions (non-additive effects) are read directly from the learned weights and nonlinear activations of a feedforward network. A minimal path criterion through a hidden unit is combined with aggregation of incoming weights ($\mu_j(I)$, often as min-abs) and output gradient bounds ($z_j$), yielding an interaction strength $z_j \cdot \mu_j(I)$. Top-ranked interactions are validated via additive models (Tsang et al., 2017).
• Topological SNI (PID): Persistent homology measures the robustness of feature subsets as topological components in the weight-induced digraph of a trained FNN, with 0-th persistence per($\mathcal{I}$) quantifying SNI strength. PID identifies interactions as those subsets whose connectivity to the output is most resistant to weight perturbation, formalized through filtration and mask matrices (Liu et al., 2020).
• Statistical–Neural Interaction for data imputation: SNI is operationalized by fusing correlation-derived statistical priors with neural feature-attention (CPFA) under controllable regularization. Each target feature’s imputation is regularized toward pairwise correlation vectors, with the trade-off governed by head-wise prior-strength coefficients $\{\lambda_h\}$. The aggregate attention map $D_{ij}$ forms a directed, interpretable model-reliance matrix, supporting transparent diagnostics of “which features influence which” beyond classical imputation (Deng et al., 18 Jan 2026).
• SNI-inspired deep learning for physics: Neural network statistical mechanics replaces the analytic Hamiltonian by a deep autoregressive density, whose negative log expresses the energy function; effective (multi-body) couplings are extracted from the learned distribution and used to recover phase diagrams and renormalization-group flow directly from configuration data (Wang et al., 2020).
• Statistical guarantees for neural networks: In shallow, linear networks (with L1 regularization), SNI principles mediate between statistical error (limited by sample size and complexity, $O(1/\sqrt{n})$) and optimization error (proximity to stationary points, $O(\varepsilon)$), guaranteeing that any approximate stationary point achieves near-optimal risk up to their sum (Taheri et al., 2022).

6. SNI in Population and Single-Neuron Inference

SNI hybridizes statistical and machine learning approaches for efficient and interpretable inference in large-scale neural population recordings:

• Population GLMs and state-space models: SNI approaches decouple time-varying stimulus effects, network interaction strengths, and latent network states by embedding pseudolikelihood or mean-field approximations (Bethe, TAP) into sequential Bayesian estimation. This enables scalable, time-resolved recovery of network entropy, sparsity, and heat capacity under dynamic, nonstationary conditions (Donner et al., 2016).
• Statistical–Neural Coding Augmentation: For receptive-field-based neuron classification, SNI leverages point-process GLMs for initial statistical labeling, augments rare classes with synthetic spike trains, and transfers these to deep convolutional classifiers for rapid labeling at scale, while maintaining interpretability of covariate effects (Sarmashghi et al., 2022).

SNI therefore represents both an analytic tool for demixing and quantifying neural dependencies, and a methodological guide for constructing robust, interpretable, and theoretically grounded models in computational neuroscience and statistical machine learning.