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 hard clipping or post-hoc asymmetric saturation with a dynamic output state that remains inside a prescribed asymmetric interval. A neural network emits a command uc, while the realized output u evolves through the APIR vector field, producing bounded actions, temporal smoothing, and gradients that remain available in the interior.
Replace a single recurrent state update with fast feature relaxation, activity evolution, and a slow adaptive state that modulates the activity vector field. Tune the activity subsystem near a controllable saddle-node so that it retains a useful transient regime for a predictable number of steps, enabling delayed switching and long-horizon memory without requiring a large hidden state.
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.
Compute an inner approximation of the states from which a neural controller can keep the plant inside a prescribed safe domain indefinitely, then use the resulting regulation map as a safety shield around the network. At each state, the network proposes an action, but the shield projects or replaces it with an action certified to remain in the invariant set.
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 single neural representation of a rapidly oscillatory PDE solution by a macroscopic network plus an explicitly oscillatory corrector network. Feed the network both the slow coordinate $x$ and fast coordinate $y=x/\varepsilon$, and train the resulting composite field in a variational energy objective. This targets the paper's scale-robust approximation bound rather than forcing the optimizer and finite sample set to discover oscillations of wavelength $\varepsilon$.
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.
Build a residual module whose state explicitly contains both a persistent context representation and an accumulator. Each residual branch computes one learned correction and adds it to the accumulator, instead of forcing every layer to represent the complete output from scratch. This provides a concrete solver-like architecture for high-dimensional regression and iterative latent prediction.
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.
Initialize each row of a neural weight matrix as a stationary correlated Gaussian process instead of using independent entries, but constrain its correlation tail to remain on the finite-fourth-moment side of the transition. This creates controllable structured spectra while avoiding the heavy-edge regime predicted for correlations slower than \(t^{-1/2}\).
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.
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.
Build a Bloch-conditioned neural model whose periodic-factor representation transforms covariantly when the supplied Bloch wavenumber is shifted by a reciprocal lattice vector. Either canonicalize q to the first Brillouin zone or augment training with mathematically paired examples whose outputs differ by the exact phase gauge. This prevents the network from learning inconsistent predictions for physically identical Bloch modes.
Replace a dense degree-m tensor interaction layer by a symmetric orbit-parameterized layer with one parameter per exponent vector and explicit multinomial scaling. This preserves the contribution of all ordered tensor entries represented by one orbit, while reducing parameter count and avoiding the amplitude distortion of unweighted monomial compression.
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.
Replace deterministic binary-tree pooling or hierarchical feature aggregation by a stochastic merge that chooses either elementwise addition or elementwise minimum. The mixing probability p controls whether zero or sparse states proliferate or disappear, with a predicted absorbing-state transition at p = 1/2.