Structured Basis Function Network

• Structured Basis Function Networks are neural architectures that represent components as expansions over fixed or adaptive basis functions with explicit algebraic, geometric, or statistical structure.
• The approach shifts learning from unconstrained weights to estimating coefficients and basis parameters, thereby imposing a predetermined hypothesis space that enhances data efficiency and stability.
• Applications span convolutional filters with Gaussian derivatives, spline-parameterized neural ODEs, and operator-learning models, all demonstrating improved interpretability and performance.

A structured basis function network denotes a neural architecture in which a target mapping, an internal representation, or a parameter field is constrained to lie in the span of basis functions with explicit algebraic, geometric, or statistical structure. Across the literature, this pattern appears in convolutional filters expanded in fixed Gaussian-derivative bases, shallow Beta basis function models, adaptive basis layers for functional data, basis-to-basis operator learners, continuous-in-depth spline-parameterized networks, structured radial basis ensemble combiners, and RBF-based Kolmogorov–Arnold networks (Jacobsen et al., 2016, Dhahri et al., 2012, Ingebrand et al., 2024, Dominguez et al., 2 Sep 2025). The common principle is that learning is shifted from unconstrained weights toward coefficients, basis parameters, or both, thereby imposing a designed hypothesis space rather than a fully free one.

1. Canonical formulation and defining characteristics

The canonical form is a basis expansion. In the Beta Basis Function Neural Network, the network output is written as

$$f(x)=y=\sum_{i=1}^{k} w_i\, B_i(x;x_{c_i},d_i,p_i,q_i),$$

where the hidden units are Beta basis functions with centers, widths, and shape parameters (Dhahri et al., 2012). In structured receptive field networks for CNNs, each convolution kernel is parameterized as

$$f^{(l)}_{m,c}(x)=\sum_{k=1}^{K_B} w^{(l)}_{m,c,k}\,\phi_k(x),$$

with a fixed basis $\{\phi_k\}$ and learned coefficients $w^{(l)}_{m,c,k}$ (Jacobsen et al., 2016). In functional data analysis, the Basis Layer computes coefficients

$$c_i=\langle \beta_i,X\rangle=\int_0^1 \beta_i(t)X(t)\,dt,$$

where each $\beta_i$ is itself a micro neural network, so the basis functions are learned while the projection structure is retained (Yao et al., 2021). In operator learning, a function is represented as

$f(x)\approx \sum_{j=1}^{k}\alpha_j g_j(x),$

and an operator is learned as a map between coefficient vectors rather than directly between sampled functions (Ingebrand et al., 2024).

This family is therefore defined less by a single topology than by a structural constraint: a network component is expressed through basis functions whose form is fixed, partially fixed, or parameterized in a low-dimensional manner. The basis may act in spatial coordinates, feature coordinates, function space, the depth coordinate of a neural ODE, or the space of ensemble predictions. What remains invariant is the replacement of an unconstrained parameter field by a structured expansion.

2. Basis families and the loci of structure

The basis family varies with the domain. In structured receptive field CNNs, the basis is a multiscale Gaussian-derivative family,

$$\phi_k(x)\equiv \partial^{\alpha_k}G_{\sigma_k}(x),$$

chosen for scale-space, multiscale, and steerability properties (Jacobsen et al., 2016). In spline-parameterized neural ODE controls, weights and biases are expanded over B-spline basis functions in depth,

$$W(t)=\sum_{l=-d}^{L-1}\omega_l B_l^d(t),\qquad b(t)=\sum_{l=-d}^{L-1}\beta_l B_l^d(t),$$

so the structure is imposed over the continuous layer index rather than over input space (Günther et al., 2021). In basis-pretrained function approximation, the basis family is polynomial, with subnetworks trained to approximate monomials or tensor-product monomials on a reference domain (Hu et al., 9 Oct 2025). In Free-RBF-KAN, each univariate component is expanded in RBFs with trainable centroids and scales,

$$g(x)=\sum_{m=1}^{G}\omega_m K\!\left(\frac{x-c_m}{\sigma_m}\right),$$

so the structure is local, adaptive, and composed according to a Kolmogorov–Arnold pattern (Chiu et al., 12 Jan 2026).

Other works move the structure to different objects. DeepLABNet replaces fixed scalar activations by per-feature RBF expansions,

$$f^{(l)}_{m,c}(x)=\sum_{k=1}^{K_B} w^{(l)}_{m,c,k}\,\phi_k(x),$$0

making the activation function itself a structured basis model (Hryniowski et al., 2019). SE-RBFNet represents a signed distance function by sparse ellipsoidal radial basis functions,

$$f^{(l)}_{m,c}(x)=\sum_{k=1}^{K_B} w^{(l)}_{m,c,k}\,\phi_k(x),$$1

where $$f^{(l)}_{m,c}(x)=\sum_{k=1}^{K_B} w^{(l)}_{m,c,k}\,\phi_k(x),$$2 includes centers, anisotropic shapes, and orientations, so the structure is explicitly geometric (Lian et al., 5 May 2025). In structured ensemble models, the basis acts on predictor outputs rather than on raw inputs: multiple hypotheses are turned into a structured dataset, and a Gaussian RBF combiner is trained on that representation (Dominguez et al., 2023).

These examples show that “basis” need not mean fixed analytic atoms in the classical approximation-theoretic sense. It may also mean neural basis modules, operator bases, localized supports over a functional domain, or basis functions defined over depth or over ensemble outputs. The structural constraint is therefore architectural and semantic, not merely algebraic.

3. Learning strategies

Training procedures are correspondingly heterogeneous. A two-level scheme appears in the hierarchical learning algorithm for BBFNN: a Genetic Algorithm determines the number of hidden units and coarse basis parameters, while gradient descent refines $$f^{(l)}_{m,c}(x)=\sum_{k=1}^{K_B} w^{(l)}_{m,c,k}\,\phi_k(x),$$3, $$f^{(l)}_{m,c}(x)=\sum_{k=1}^{K_B} w^{(l)}_{m,c,k}\,\phi_k(x),$$4, $$f^{(l)}_{m,c}(x)=\sum_{k=1}^{K_B} w^{(l)}_{m,c,k}\,\phi_k(x),$$5, $$f^{(l)}_{m,c}(x)=\sum_{k=1}^{K_B} w^{(l)}_{m,c,k}\,\phi_k(x),$$6, and $$f^{(l)}_{m,c}(x)=\sum_{k=1}^{K_B} w^{(l)}_{m,c,k}\,\phi_k(x),$$7 using explicit derivative updates (Dhahri et al., 2012). In structured receptive field CNNs, only the combination weights are trainable; the basis convolutions are fixed, and the layer reduces to basis responses followed by learned linear combinations (Jacobsen et al., 2016). In basis-to-basis operator learning, coefficients are computed by least squares, and the operator map is then learned in coefficient space; for linear operators this becomes a single matrix fit with a closed-form solution (Ingebrand et al., 2024).

Closed-form or partially closed-form training recurs in ensemble settings. In the Structured Radial Basis Function Network for multiple-hypothesis regression, predictor outputs define centers and scales for RBF units, and the output-layer weights are obtained by ridge-style least squares (Dominguez et al., 2023). The later loss-centric structured basis function framework generalizes this viewpoint by making the combiner consistent with the geometry of the loss through Bregman-centroid aggregation and providing both a closed-form least-squares estimator and a gradient-based procedure (Dominguez et al., 2 Sep 2025). In polynomial basis pretraining, basis networks are first trained offline on a reference domain, then reused through domain mapping and linear recombination, with coefficients obtainable by least squares or further fine-tuning (Hu et al., 9 Oct 2025).

Bayesian and stateful variants introduce additional mechanisms. In sparse Bayesian deep functional learning, a group spike-and-slab prior is placed on first-layer columns associated with localized B-spline features, and MAP estimation is combined with posterior inclusion probabilities for region selection (Zhu et al., 24 Feb 2026). In stateful ODE-Nets using basis expansions, both trainable parameters and state parameters are represented continuously in depth; state updates are projected back onto the basis by a least-squares regression over sampled time points (Queiruga et al., 2021). This suggests that structured basis function networks often separate representation learning, coefficient estimation, and structural adaptation rather than collapsing all parameters into a single unconstrained backpropagation problem.

4. Theoretical properties and inductive bias

The principal theoretical role of structure is to restrict the hypothesis class in a controlled way. In spline-parameterized neural controls, decoupling trainable parameters from the number of layers allows accuracy of network propagation to be studied separately from the optimization problem, and the reported effect is increased robustness of the learning problem toward hyperparameters due to increased stability and accuracy of the network propagation (Günther et al., 2021). In continuous-depth basis-parameterized ODE-Nets, higher-order integrators induce an implicit regularization term involving derivatives of the underlying dynamics, which the paper interprets as supporting compressibility of the learned depth-dependent weights (Queiruga et al., 2021).

Approximation theorems are explicit in several instances. AdaFNN proves consistency for maps of the form $$f^{(l)}_{m,c}(x)=\sum_{k=1}^{K_B} w^{(l)}_{m,c,k}\,\phi_k(x),$$8, where $$f^{(l)}_{m,c}(x)=\sum_{k=1}^{K_B} w^{(l)}_{m,c,k}\,\phi_k(x),$$9 is a linear continuous map on $\{\phi_k\}$0 and $\{\phi_k\}$1 is a continuous map on a finite-dimensional coefficient space, and also derives a small-generalization-error result under boundedness and Lipschitz assumptions (Yao et al., 2021). Free-RBF-KAN proves a universality result: if $\{\phi_k\}$2 is continuous and nonpolynomial, then the span

$\{\phi_k\}$3

is dense in $\{\phi_k\}$4, and the resulting NP-KAN and RBF-KAN architectures are universal approximators on $\{\phi_k\}$5 (Chiu et al., 12 Jan 2026). Sparse Bayesian deep functional learning establishes approximation error bounds, posterior consistency, and region selection consistency for its B-spline-based functional embedding plus deep Bayesian network (Zhu et al., 24 Feb 2026).

Operator-learning formulations emphasize a different kind of rigor. Basis-to-Basis operator learning separates basis learning from coefficient mapping, exploits Hilbert-space projection through least squares, and for linear operators obtains a coefficient-space matrix transformation with a closed-form solution; the same framework yields analogues of eigendecomposition and singular value decomposition (Ingebrand et al., 2024). A plausible implication is that structured basis function networks frequently gain interpretability not from post hoc attribution, but from the explicit semantics of coefficients, basis supports, or spectral factors.

5. Architectural variants and application domains

The framework is not confined to any single modality. In computer vision, structured receptive field networks use Gaussian-derivative filter bases and report improvements over unstructured CNNs in small and medium dataset scenarios, state-of-the-art classification results on popular 3D MRI brain-disease datasets, and competitive behavior on larger datasets such as ILSVRC2012 (Jacobsen et al., 2016). DeepLABNet embeds learnable radial-basis activations inside standard deep architectures and evaluates them on MNIST, CIFAR-10, and CIFAR-100, including ResNet-20 and larger ResNets (Hryniowski et al., 2019). In graphics and geometry processing, SE-RBFNet fits sparse ellipsoidal radial basis functions to signed distance fields of point clouds, using an octree-based hierarchy, nearest-neighbor locality restriction, and CUDA parallelization (Lian et al., 5 May 2025).

For functional and operator-valued data, the architecture becomes explicitly infinite-dimensional before truncation. AdaFNN learns basis functions as micro networks and applies them to regression and classification with functional inputs (Yao et al., 2021). sBayFDNN builds B-spline projections of functions, applies a deep Bayesian network to the resulting basis coefficients, and interprets first-layer column sparsity as region selection over the functional domain (Zhu et al., 24 Feb 2026). B2B operator learning works on Hilbert spaces of functions and reports validation on seven benchmark operator learning tasks, with a two-orders-of-magnitude improvement in accuracy over existing approaches on several benchmark tasks (Ingebrand et al., 2024). Free-RBF-KAN is evaluated on multiscale function approximation, physics-informed learning for heat and Helmholtz equations, and DeepONet trunk design for reaction–diffusion operator learning (Chiu et al., 12 Jan 2026).

Ensemble and uncertainty-oriented variants place the basis in prediction space. Structured RBF ensembles for regression use multiple hypotheses, Voronoi tessellations, and an RBF combiner on the structured prediction matrix, with empirical studies on air quality and energy appliance prediction (Dominguez et al., 2023). The later loss-centric s-BFN framework extends this to regression and classification under Bregman geometry, with a tunable diversity parameter $\{\phi_k\}$6 controlling specialization and aggregation behavior (Dominguez et al., 2 Sep 2025). This broad dispersion of applications indicates that the defining property is not the domain but the existence of an explicitly designed basis representation inside the learning system.

6. Limitations, misconceptions, and open directions

A common misconception is that structured basis function networks are synonymous with classical shallow RBF networks. The literature contradicts that identification: the basis may parameterize CNN filters, scalar activations, functional projections, continuous-depth controls, operator encoders, or ensemble combiners, and it may be fixed, partially trainable, or fully adaptive (Jacobsen et al., 2016, Hryniowski et al., 2019, Günther et al., 2021). Another misconception is that structure implies rigidity comparable to a fully engineered representation. Several models instead learn arbitrary effective filters from a fixed basis, train activation centroids and smoothness, or fine-tune pretrained basis modules end-to-end (Jacobsen et al., 2016, Chiu et al., 12 Jan 2026, Hu et al., 9 Oct 2025).

The trade-offs are equally recurrent. Restricting the span of admissible functions can improve data efficiency and stability, but it can also underfit if the chosen basis is poorly matched to the task. This limitation is stated directly for structured receptive field CNNs, for continuous-depth spline controls with too few coefficients, and for B2B operator learning when the learned bases do not approximate the underlying function spaces well (Jacobsen et al., 2016, Günther et al., 2021, Ingebrand et al., 2024). Polynomial basis libraries face the $\{\phi_k\}$7 growth of monomials in higher dimension, which the authors identify as a key scalability challenge (Hu et al., 9 Oct 2025). Deep RBF activation models require careful initialization, regularization, and clipping to avoid instability (Hryniowski et al., 2019). Sparse geometric RBF models require explicit strategies for addition, elimination, and neighborhood restriction to remain computationally feasible (Lian et al., 5 May 2025). Loss-centric multi-hypothesis ensembles require tuning of diversity, since too little diversity produces collapse while too much can reduce generalization (Dominguez et al., 2 Sep 2025).

Current directions therefore concentrate on adaptive basis design, coefficient-space learning, basis compression, and uncertainty-aware selection. The evidence across these works suggests that a structured basis function network is best understood not as a single architecture but as a design paradigm: impose a basis with mathematically meaningful structure, learn coefficients and selected basis parameters in that structured space, and exploit the resulting inductive bias where fully free parameterizations are statistically, computationally, or semantically inefficient.