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.
Estimate the irreducible rejection caused by missing within-block information, then use it to choose the block's conditioning order instead of tuning block length blindly. If the estimated floor is high, expose one or more realized tokens before continuing; if the floor is low but observed rejection is high, spend compute on improving the drafter.
Use a spherical-design codebook and the paper's polar slack factorization to create a nonnegative geometric interaction bias for attention or expert routing. The resulting kernel is generated by a rank-one term and a rank-at-most-d term, and entries close to zero can define a structured sparse mask instead of relying only on learned top-k selection.
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 random projection used before retrieval, classification, or expert routing with a publicly reproducible matrix generated by a Pisot beta-transformation orbit. Search over a small public seed and sampling gap to select one matrix that preserves the calibration set's pairwise distances, then freeze it for training and inference. The projection removes random-matrix storage and makes the same embedding transform exactly reproducible across servers or proof systems.
Evaluate weak residuals against a bank of periodic trigonometric test functions using FFT projections instead of repeated pointwise quadrature or output automatic differentiation. Frequency truncation and mode weighting provide a direct way to control the spatial scales enforced during neural PDE 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.
Replace a large diagonalizable recurrent or state-space transition operator by a sparse set of retained oscillatory modes selected according to their contribution to the output autocorrelation. Unlike magnitude-based pruning, the objective is to preserve the power-law return signal generated by pairwise spectral differences, enabling long memory with far fewer modes.
Use the KS ratio to decide how many message-passing layers to execute per graph or per node, rather than selecting a fixed depth. In the subcritical regime, stop once the predicted remaining effect is below a tolerance; in the supercritical regime, continue until the observed logit change becomes small or a larger budget is reached.
Replace dense self-attention by a sparse mask on a one-dimensional or ordered token geometry, retaining local neighbors and adding long-range edges with probability proportional to distance raised to \(-(1+\sigma)\). Use \(\sigma\approx0.8\text{--}0.85\) as the initial regime because the paper finds that this range supports delocalized, GOE-like connectivity despite sparse bonds. The resulting layer has linear or near-linear attention cost while maintaining long-range paths.
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.
Train a residual network on a coarse depth mesh, estimate a dual-weighted residual for every layer interval, and insert new layers at intervals with the largest estimated contribution to objective error. This replaces uniform depth expansion or expensive neural architecture search with targeted refinement driven by both forward-dynamics error and downstream loss sensitivity.
Construct a sparse attention layer by sampling backward token histories as a continuous-time branching process rather than allowing every query to attend to every key. Each active ancestor either dies or branches into a bounded number of candidate ancestors, with branching probability controlled by a small parameter. The branch-out penalty predicts exponentially small probability of long, highly branching histories, providing a direct knob for receptive-field size and attention FLOPs.
Train a network to predict the context-dependent observation matrix rather than the latent inverse parameters themselves, then compute the latent parameters with a differentiable ridge-regression solve. This gives one model that can assimilate arbitrary observation vectors, exposes the conditioning of the inverse problem, and avoids forcing an MLP to learn the entire map from observations to parameters.
Replace a dense coordinate-kernel interaction among N points by an orthogonal samplet transform with a sparse detail-detail matrix and a small polynomial branch. Detail basis vectors have vanishing moments, so smooth low-frequency behavior is represented by a few polynomial coefficients while localized residual interactions become sparse in the transformed domain.
Replace independent top-k expert selection with a decision-diagram router that optimizes a quadratic surrogate over binary expert activations and continuous assignment weights. When expert redundancy or communication costs form a tree, partial routing decisions are merged whenever they have the same separator state, turning exponentially many candidate masks into a dynamic program over a small number of graph cuts.
Replace a large dense layer whose input and output dimensions factor into multiple modes by a TT-matrix whose parameters are stored as a chain of small cores. Periodically apply TT-SVD rounding to remove weak singular directions and keep the representation within a prescribed approximation error. This transfers the paper's central computational principle—perform tensor-product contractions directly in compressed form—to neural network layers.
Replace a dense weight matrix by a cross approximation built from selected rows and columns rather than by a conventional truncated SVD. Periodically refresh the selected indices using residual leverage scores, warm-starting from the previous factorization so that the compressed layer can track weight changes during fine-tuning.
Represent every mesh interface degree of freedom by one feature copy per incident cell, and apply local neural blocks directly to these cell tensors. Enforce inter-cell consistency with valence-weighted averaging only after selected layers or hierarchy transitions, avoiding repeated construction of a global sparse graph or assembled feature vector. This is suited to adaptive quadtrees, octrees, and finite-element neural operators.
Use the expander decoder as a structured sparse-coding dictionary and replace dense OMP correlation steps with edge-wise gather-and-reduce operations. This is useful when codes must be inferred iteratively, including interpretable feature extraction, sparse retrieval, or an inference-time latent selector that cannot rely entirely on an amortized encoder.
Replace ordinary masked mean pooling with a Fourier-compressed quadrature operator for arbitrary two-dimensional or three-dimensional domains. The geometry is preprocessed once into reusable grid weights, allowing every channel and every training example using the same domain to be pooled without boundary-area bias.
Replace dense grid tokens or global spectral features with coefficients of compactly supported kernels centered on a nested hierarchy of spatial points. Encode an input field into coarse-to-fine coefficients, apply a neural map to those coefficients, and decode the predicted coefficients at arbitrary query locations; the contribution from each level provides an explicit multiscale output decomposition.