Attention variants, state-space and recurrent cells, normalization and token-mixing schemes — each tested against the standard block it replaces.
Replace Cox-de Boor evaluation of each cubic B-spline edge activation with its fixed truncated-power expansion. Normalize each scalar edge input to a bounded knot coordinate, evaluate the five shifted cubic positive-part terms in parallel, and contract them with the learned spline coefficients inside one fused kernel.
Construct a residual network from independently attachable modules, but permit only a number of modules whose aggregate feedback gain lies inside a delay-dependent admissible interval. Estimate deployed end-to-end latency and each module's local Jacobian gain, then reject or bypass additional modules when the predicted delayed-loop stability boundary is crossed. This turns variable-width or depth scaling into a falsifiable control problem rather than an empirical choice.
Replace black-box differentiation through an embedded LP decision with an analytic Jacobian computed from the LP’s active basis. A neural policy emits LP coefficients or right-hand sides; the LP returns the decision, while the backward pass uses the basis inverse and dual sensitivity, avoiding solver unrolling and finite-difference noise.
Replace a square dense projection in a Transformer or MLP with a trainable recursive butterfly matrix. The layer preserves multiscale channel interactions while constraining every complementary row-column block to rank at most k, reducing parameters and enabling recursive structured matrix-vector products. Unlike an arbitrary sparse layer, the construction has an explicit recursive factorization and a quasi-optimal approximation guarantee among matrices with the same butterfly rank.
Replace a dense neural-network weight tensor with a graph tensor network whose physical modes and internal edge ranks are specified by a sparse rank-adjacency matrix. Unlike tensor-train or hierarchical Tucker layers, the graph can contain selected cycles and skip connections between tensor modes, allowing the factorization topology to match correlations in the weight tensor. Fit the layer with GTN-SVD at a prescribed tolerance and compare accuracy, parameter count, and tensor-contraction…
Compress each hidden layer by retaining directions that are simultaneously reachable from the observed input distribution and observable at the network output. Unlike PCA or SVD, the retained subspace is weighted by downstream task sensitivity, so high-variance but output-irrelevant directions can be removed while low-variance predictive directions are preserved.
Represent a large linear map acting on a Cartesian 3D grid and multiple physical channels as a TT-matrix, while retaining separate TT blocks for channel couplings that have different semantics. Apply the layer by sequential contractions with TT cores rather than materializing a dense matrix or a full 3D convolution kernel. Rank truncation provides an explicit accuracy-versus-memory knob and can be applied after optimizer updates.
Replace a trainable shallow MLP hidden layer by a frozen bank of smooth sigmoid ridge functions and train only a linear output head. Choose the feature count and parameter sampling regime using the theorem's explicit dependence on input dimension d, target regularity k, evaluation norm m, and confidence delta. The construction is especially appropriate for smooth regression, scientific surrogate models, and PINNs, where derivatives of the network output are part of the loss.
Replace a dense Haar or Gaussian random projection with a streamed product of random two-coordinate rotations followed by coordinate subsampling. The transform is exactly orthogonal before subsampling, requires only a list of rotation triples, and the paper's pseudo-mixing result predicts that degree-two statistics relevant to norm preservation and Johnson–Lindenstrauss embeddings become Haar-like after only O(n polylog(n)) rotations.
Insert a linear Johnson–Lindenstrauss bottleneck around a set of jointly processed representations, choosing its width from the sharp finite-set dimension bound rather than from the model's nominal hidden size. The projection should preserve pairwise distances between tokens, patches, or retrieved items, allowing a downstream attention or MLP block to operate at lower width while retaining the geometry relevant to similarity computations.
Build a neural operator around explicit input and output measurement spaces rather than forcing the network to consume and emit a fixed grid. The same learned latent surrogate can be reused on alternative sensor layouts or query meshes through reconstruction and re-encoding maps, with a consistency loss enforcing agreement between measurement pipelines.
Attach a polyharmonic spline decoder to a coordinate MLP or use it as a standalone neural-field output head over a large set of spatial anchors. The decoder represents the output as a low-degree polynomial trend plus a PHS kernel expansion, while FMM evaluates all anchor-to-query interactions in approximately linear or near-linear cost. When coefficients must be fitted or periodically recalibrated, solve the constrained interpolation system with projected conjugate gradients and a sparse…
Replace dense continuous action search during neural-controller verification with a finite set of representative inputs induced by affine pieces of the interval neural dynamics. This makes safety checking parallel over state cells and candidate actions, enabling much cheaper certification or repeated safe-set updates.
Replace fixed-length binary dot products with accumulations whose terms are processed in descending order of weight magnitude. Stop as soon as the current partial sum is larger in magnitude than the total absolute magnitude of all remaining terms; the output sign is then guaranteed to equal the full dot-product sign, eliminating unnecessary additions without changing accuracy.
Build an SPD classifier and residual head directly from Cholesky factors, using lower-triangular differences and matrix-power terms instead of generic eigendecomposition-based logarithm operators. This retains covariance geometry while making positive-definiteness automatic and backpropagation more numerically stable for minibatch training.
Replace Lorentz-hyperboloid tensors with proper-velocity tensors whose spatial coordinates can be transformed by standard Euclidean affine layers and activations. Reconstruct the Lorentz time coordinate only at manifold boundaries, preserving the hyperbolic representation while avoiding repeated projection, normalization, or fragile exponential-map calculations.
Grow a neural network by appending a trainable block together with an analytically initialized inverse block, so the newly added depth is exactly the identity at insertion time. After insertion, untie and optimize the two blocks independently; this preserves the current function while providing additional trainable degrees of freedom. For architectures with one expensive mixing operation followed by cheap channelwise blocks, the same construction can increase depth without repeatedly paying for…
Convert a sequential modular network into parallel execution layers by placing mutually commuting operators in the same layer. The resulting circuit preserves all noncommuting precedence constraints while exposing safe concurrency and fusion opportunities for inference or training.
Replace an expensive proximal activation or implicit optimization layer with a Gaussian barycentric estimator computed from energy evaluations. The resulting map is smooth and has a provable cocoercivity guarantee when the energy is weakly convex, making it a stable alternative to unconstrained learned activations or iterative proximal solvers.
Replace a weight-tied residual or neural-ODE stepper with an explicit Runge–Kutta method satisfying the reused-last-stage conditions. The final derivative is evaluated at the exact endpoint and becomes the first derivative of the next step, saving one expensive neural-vector-field call per step while preserving the designed integration order.
Represent a high-order feature tensor as a tensor train and replace a dense global feature transform by a truncated polynomial in a learned nearest-neighbor operator. The block computes a short Krylov expansion, p_m(A)x = sum from k=0 to m of c_k A^k x, compressing back to a fixed TT rank after each operator application; locality is intended to prevent rank growth from scaling with the total number of tensor sites.
Use the paper's finite-width O(n^{-1/2}) Gaussian-process approximation bound as a width-budgeting rule rather than choosing every hidden dimension uniformly. Estimate an architecture-specific constant for each layer or attention contraction, then allocate width according to the smallest dimension satisfying its allowed distributional error. This should produce narrower models at comparable GP-like behavior, or permit the same parameter budget to be concentrated in the layers where finite-width…
Use cumulative control measures as the input to a neural surrogate instead of the full sequence of control values. For a quasi-static free-boundary system satisfying the paper's average-pressure path-independence assumption, two nonnegative control histories with identical integrals up to time t should produce the same state at t, allowing a smaller training input and fewer distinct control trajectories in the dataset.
Construct a sparse fixed orthogonal mixer by repeatedly applying pi/4 rotations to randomly matched pairs of feature coordinates. Place this mixer before top-k feature pruning, sparse projection, or activation quantization so that information is spread across coordinates without using a dense random matrix.