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 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.
Constrain a student policy to transform its action in the same way that the input state is transformed, while constraining its value estimate to remain unchanged. During distillation, augment every teacher-student pair with several symmetry-transformed copies and penalize disagreement after transforming the student action back to the original frame.
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.
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.
Replace the shared hidden trunk of a multi-output regression network with a small bank of differentiable symbolic units, then let every output use a sparse additive or multiplicative combination of the same units. The architecture explicitly tests whether outputs share a latent mechanism instead of merely sharing arbitrary neural features, improving identifiability and producing equations that can be inspected or exported.
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 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 the dense decoder of an overcomplete activation autoencoder with a fixed left-d-regular expander mask and learn only the nonzero decoder values. Tie the encoder to the transpose of this sparse decoder, preserving the activation dimension m, latent width n, and TopK sparsity k while reducing learned dictionary parameters from mn to dn.
Build a sequence or graph module whose state is a canonical set of affine tropical pieces rather than an opaque hidden vector. Compose consecutive segments by Minkowski addition of their lifted supports, merge alternatives by union followed by lower-hull reduction, and evaluate the resulting piecewise-linear function with a minimum.
Replace an opaque MLP vector field with a stack of trainable symbolic primitives that can express linear terms, monomials, products, and related analytic operations. Apply an L1 penalty and prune small primitive coefficients after rollout training, yielding a compact dynamics module that is cheaper to evaluate and easier to inspect.
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.
Replace the first dense layer on q-ary categorical features by a Fourier interaction layer containing only monomials whose coordinate support is at most s. Use a Bohnenblust–Hille-inspired quasi-norm on coefficients, separately for each interaction order, to prevent a small number of high-order interactions from dominating the output. The resulting model has an explicit interaction-order knob and can be tested against a dense MLP at matched parameter count.
Replace an unconstrained scalar activation or nonnegative gate with a finite positive mixture of rational Bernstein basis functions. The learned function is monotone and concave on the nonnegative half-line, while its derivatives have controlled alternating signs; this can prevent pathological feature amplification and gives an interpretable shape prior. Use the paper's sharp exponent restriction τ≤1/2 rather than treating the power as an arbitrary hyperparameter.
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.
Replace a dense channel or token-mixing matrix with a product of positive bidiagonal factors, so information propagates through a controlled sequence of local couplings rather than arbitrary signed interactions. Initialize the factors from the paper's barycentric-subdivision factorization, then learn positive diagonal and off-diagonal parameters; the resulting map is structured, parameter-efficient, and constrained to remain totally positive.
Replace scalar neural activations by pairs of nonnegative channels whose ratio represents the signed or unsigned activation. Implement multiplication and addition through pair algebra, and renormalize each pair because the representation is invariant under multiplying both rails by the same positive scalar. This creates an explicitly bounded, cancellation-aware arithmetic layer for deep multiplicative MLPs, rational networks, and neural fields.
Replace an unconstrained high-order polynomial interaction module with features generated by Gaussian matrix contractions and their exact Wick expansion. The resulting interactions are sums of products of power-sum invariants, with coefficients fixed by perfect-matching counts, providing a low-parameter inductive bias for permutation- or orthogonal-structured data.
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.