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.
Replace selected residual, recurrent, or state-space blocks by modules whose input-output Jacobians satisfy an IODP inequality throughout a prescribed activation domain. The constraint controls incremental amplification between two trajectories without requiring either trajectory to remain near one fixed equilibrium, so it should improve robustness to changing contexts and prevent exploding long-horizon sensitivities.
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 standard ReLU surrogate with an input convex neural network whose hidden-to-hidden weights are constrained to be nonnegative. The network remains piecewise linear and expressive, but its convexity allows downstream minimization to use continuous ReLU epigraph constraints instead of binary activation variables, potentially eliminating the integrality bottleneck of neural optimization.
Represent each neural module as a Hamiltonian storage system and connect modules through a state-dependent skew or Dirac interconnection instead of arbitrary residual additions. The coupling may change with the hidden state, but its internal power contribution cancels exactly, so total stored energy is controlled only by external inputs and explicitly added dissipation.
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…
Replace direct learning of a highly cancelling signed observable with a quotient-space model over symmetry orbits of inputs. Predict a physically constrained baseline for each family and use an LSTM or set/graph encoder only for the residual many-body correlation, then aggregate family predictions with known signed weights instead of forming a noisy sample-level ratio.
Partition a neural network into independently trained or independently monitored modules and constrain their cross-module interaction gain using a compositional contraction certificate. This enables stable deep modular MLPs, graph blocks, or recurrent modules without estimating the full network Jacobian, while providing an explicit coupling threshold for when the architecture loses contraction.
Replace a deep feed-forward block by the fixed point z=phi(Wz+Vx+b), with the recurrent weight W constrained so that the fixed point is unique for every input. The same condition makes forward fixed-point iteration stable and makes implicit differentiation well-conditioned, allowing depth-independent memory usage while providing a measurable spectral failure boundary.
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.
Construct a continuous-depth or recurrent residual block with a prescribed polynomial Lyapunov decay near its equilibrium. The architecture combines a fixed radial stabilizer with a learned component that is constrained to have zero radial projection, allowing slow algebraic transients and long memory while preventing asymptotic hidden-state growth.
Replace the usual isotropic Gaussian random Fourier features with a frequency distribution matched to the expected spectral regularity of the target function. For coordinate fields, operator-learning maps, or PDE solution surrogates, this should place more features where the target Fourier energy lies and improve approximation at the same feature count. Stabilize the resulting feature matrix with whitening or ridge regression because spectral accuracy can create severe ill-conditioning.
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 single smooth neural vector field with a finite collection of smooth subnetworks selected by learned affine hyperplanes. The architecture exposes switching geometry directly, allowing it to represent friction-like or threshold dynamics without approximating discontinuities using excessively steep activations.
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.
Make a neural network predict a positive Gaussian-mixture representation of the distribution function rather than independent values on a momentum grid. Use the mixture parameters inside a differentiable Boltzmann collision operator, so training directly enforces the interaction mechanism and exposes the relaxation spectrum responsible for ballistic-to-hydrodynamic crossover.
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.
Replace a wide collection of interchangeable near-zero branches with a module whose output is explicitly a quadratic form in the branch-weight Gram matrix. The module preserves the paper's leading-order behavior while making the relevant collective variable explicit and allowing low-rank parameterizations.
Replace a single residual stream or unconstrained hyper-connection with S parallel feature streams whose cross-stream mixing matrix is doubly stochastic. Parameterize the matrix with Sinkhorn normalization so every layer preserves total stream mass while still learning adaptive information routing. This is a low-overhead alternative to dense cross-stream attention and should reduce stream explosion, collapse, and sensitivity to depth.
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 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.