Attention variants, state-space and recurrent cells, normalization and token-mixing schemes — each tested against the standard block it replaces.
Replace dense attention over structured object tokens with attention over role-filler tensor-product representations. A learned query specifies both a role and a filler, retrieves objects matching that binding, extracts a target role, and rebinds the extracted filler into an output object.
Replace the usual sum of pairwise modality similarities with a higher-order score based on the coordinatewise Hadamard product of all normalized modality embeddings. For modalities indexed by i=1,...,m, score a tuple using s(x_1,...,x_m)=\omega^\top(\bar g_1(x_1)\odot\cdots\odot\bar g_m(x_m)), where \omega is learned and \odot is coordinatewise multiplication. This adds explicit m-way interactions without concatenating raw features or introducing a joint encoder.
Replace uniformly spaced history taps in a neural state-space encoder with a fixed or learned set of non-uniform delays. Regularize the resulting delay-observation matrix to have a large smallest singular value, which makes latent-state reconstruction less sensitive to irregular timestamps and observation noise. This is directly applicable to event-based data, missing timestamps, and systems with multiple time scales.
Constrain the first convolutional layer, or every convolutional layer, by averaging each kernel over the 48 rotations and reflections of the cubic point group. A scalar 3D field then receives exactly the same prediction after any lattice rotation or reflection, eliminating the need to learn equivalent crystallographic orientations from separate examples.
Replace a fragile learned similarity score for continuous object coordinates with a projection residual against a learned signal subspace. Candidate coordinates are represented by normalized Fourier or positional feature vectors, and the score is the fraction of feature energy outside the estimated subspace. The score remains useful even when the estimated subspace is adversarially rotated, because the perturbation is controlled directly by a sine-theta distance rather than by assumptions about…
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.
Use several phase-locked states as distinct attractors of one recurrent network and shape their basin asymmetry through the phase-delay composite spectrum. This creates a controllable associative-memory architecture in which a desired memory receives a larger basin without adding a separate classifier or explicit nearest-neighbor lookup.
Build a neural operator from frozen ambient mechanism blocks and a geometry-specific algebraic constraint adapter. The adapter parameterizes all outputs in the affine set satisfying sampled linear constraints exactly, so the network never produces boundary-violating states and does not require a penalty coefficient or post-step projection.
Replace a conventional linear decoder in an autoencoder or latent state-space model with an explicit quadratic manifold decoder, allowing a small latent vector to represent curved and transport-like state trajectories. Add a dynamics-aware invariance loss that penalizes the discrepancy between the time derivative of the quadratic manifold and the neural dynamics evaluated on that manifold.
Replace fixed random Fourier features in a coordinate network or PINN with a small complex Fourier dictionary whose propagation directions are learned from a weighted residual. Given directions, solve the linear feature coefficients exactly or by ridge regression, and optimize only the directions in the outer loop. This should represent low-directional-complexity fields with fewer features and avoid wasting gradient updates on coefficients that can be fitted analytically.
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.
Parameterize an entropic OT cost only in directions that can change the transport plan, removing row-plus-column potential directions that are invisible because of OT gauge invariance. Whiten the remaining feature coordinates using their empirical covariance, producing an OT layer whose identifiable parameters have substantially more uniform sensitivity.
Replace an oversized recurrent hidden state or raw history stack with a causal filtered input-output lift followed by an SVD-selected bottleneck. The actor, critic, and Bellman regression operate only on the identifiable memory coordinates, preventing deterministic null directions from being fitted as if they were independent state variables.
Freeze a randomly initialized single-layer transformer and use a constructed soft prompt to make its attention weights equal Gaussian-kernel weights over support examples. The resulting model performs Nadaraya-Watson regression in one forward pass, so task adaptation stores prompt tokens rather than modifying network weights. Prompt length becomes the number of kernel centers, while hidden dimension and prompt norm determine whether the required logits can be represented accurately.
Replace an unconstrained edge-feature residual update in a graph neural network with separate cut-space and harmonic-space updates. The cut branch carries transfer information visible at nodes, while the harmonic branch models cycle circulation and can be given an independently chosen contraction rate, preventing persistent or unstable circulation features from contaminating node predictions.
Use predicted covariance reduction as a differentiable gate for selecting tokens, views, sensors, or retrieved demonstrations. The gate favors inputs with high expected information gain while accounting for acquisition cost, turning attention and data collection into active observability optimization.
Use a bank of damped rotational state channels with a deliberately spread decay spectrum, allowing one recurrent layer to represent short, medium, and long temporal dependencies without relying on a single learned spectral radius. Concatenate the channels and train a readout or downstream nonlinear head to select the appropriate memory timescale.
Replace an unconstrained recurrent matrix with an orthogonally mixed block diagonal matrix whose blocks are independently parameterized damped rotations. The model receives explicit phase mixing from the rotation frequencies and controlled forgetting from the decay rates, while its linear recurrent dynamics have a known contraction factor before the nonlinear activation.
Build a complex-valued recurrent or graph-neural layer whose hidden state evolves under a fixed graph Schrödinger operator and is exposed to the downstream network only through coordinate magnitudes at several times. Choose the diagonal potential so that the spectrum has unique unordered pair sums, the squared-eigenvector matrix is invertible, and every eigenvector pair overlaps in at least one observed coordinate; the resulting temporal intensity code is theoretically injective up to one…
Replace Gaussian covariance propagation in a neural state-space model with a finite Perron–Frobenius operator acting on coefficients of a learned density basis. A neural encoder maps observations to latent states, while an eDMD-derived matrix transports the full coefficient vector and supports multimodal or skewed uncertainty. This creates a cheap deterministic uncertainty layer that can be rolled forward for long horizons without repeatedly sampling particles.
Replace independently parameterized scalar, vector, and higher-order neural outputs with consecutive spaces of ReLU-power differential forms linked by an exact exterior-derivative layer. The network can then produce curl-free, divergence-free, or more general closed fields by construction, while the complex prevents artificial null-space modes that commonly appear when differential constraints are enforced only through sampled residual losses.
Replace a portion of quadratic key-value attention or an external episodic table with a per-sample matrix fast memory updated by rank-one delta corrections. The memory directly learns a linear key-to-value map and can be carried across sequence segments, providing cheap online adaptation with constant state size per head.
Add a differentiable orientation-pooling layer after steerable filters or an orientation-bin expansion. It aggregates unoriented line evidence using doubled-angle vectors, so a feature at angle θ is identical to one at θ+π, while symmetric orientations cancel naturally instead of producing an arbitrary mean angle. Feed the network both the Cartesian order parameter and its magnitude-based confidence.
Replace raw molecular orientation vectors with local scalar features invariant under common three-dimensional rotations and the apolar transformation u_i -> -u_i. Feed these channels to a CNN autoencoder, VAE, or contrastive encoder so that configurations on the same physical symmetry orbit have identical inputs or latent codes. This should improve unsupervised phase discovery without supplying order-parameter labels.