Machine-learning ideas tagged Initialization in the ML taxonomy of the Math2NN corpus.
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 uniform collocation for a fixed random-feature neural PDE solver with sampling from the leverage-score density of the operator-applied features. Whiten the retained residual feature space before solving for output coefficients, so the sampled least-squares matrix has an identity-like expected Gram rather than inheriting severe anisotropy from the differential operator. The same construction can be used for a linearized neural network by treating Jacobian features as the trial functions.
Train two parameter replicas with symmetric coupling, treating one replica as a prepared thermalization packet for the other. Estimate the slow local Hessian direction and initialize or periodically reset the packet so that the coupled state has zero projection onto that mode; the target should then relax according to the next-slowest mode rather than the original bottleneck.
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.
Choose an initialization that may have worse initial loss but has a smaller projection onto the slow modes of the subsequent training dynamics. Under the same optimizer, data order, and learning rate, this initialization should overtake a lower-loss baseline after a predictable crossing time, analogous to the paper's reversal of relaxation ordering.
Grow a neural network by appending a trainable block together with an analytically initialized inverse block, so the newly added depth is exactly the identity at insertion time. After insertion, untie and optimize the two blocks independently; this preserves the current function while providing additional trainable degrees of freedom. For architectures with one expensive mixing operation followed by cheap channelwise blocks, the same construction can increase depth without repeatedly paying for…
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.
Replace ordinary momentum with a semi-implicit velocity update containing viscous damping and a proximal dry-friction step, while evaluating the gradient at a look-ahead parameter point. The dry-friction proximal operator exactly zeros sufficiently small velocities, which may suppress late-training oscillations and create finite-time stationarity instead of merely asymptotic velocity decay.
Normalize a learned symplectic recurrent transition using the paper's explicit positive-definite factor instead of projecting the entire matrix onto an orthogonal group. Retain the compact unitary dynamics as the stable transport component and separately damp only the noncompact factor, providing a tunable stability mechanism that preserves symplectic structure.
Split the optimizer state into a passive descent sector and an explicitly nonequilibrium current-like sector, then constrain the local update Jacobian to be block triangular. The active sector may perturb the passive update, but passive perturbations cannot feed back into the active sector at first order, making its decay rate measurable and preventing uncontrolled oscillatory feedback. A second design objective tunes the two decay rates to reproduce the paper's slow-mode splitting and uses the…
Parameterize an orthogonal or semi-orthogonal neural weight matrix directly on the Stiefel manifold and update it with a Cayley retraction instead of unconstrained SGD plus a penalty or QR projection. The update preserves orthogonality exactly, is second-order accurate for the appropriate metric, and avoids the cubic QR factorization at every optimizer step.
Replace projected or retracted updates for constrained spherical parameter blocks with a geodesic Frank–Wolfe update and an iteration-only step size \(\eta_k=a/(k+a)\). The method moves along a minimizing geodesic toward a feasible linear-oracle point, avoiding repeated projection and eliminating line-search or gap-feedback overhead. On locally error-bounded objectives, the paper predicts accelerated polynomial convergence, including \(O(k^{-2})\) for strongly geodesically convex objectives…
Use paired recurrent channels with exactly reciprocal gains while applying a common phase rotation. One channel carries a controlled expanding mode and the other a matching contracting mode, creating a tunable hyperbolic memory spectrum without the optimization fragility of an unconstrained recurrent matrix.
Replace a trainable shallow hidden layer by a deterministic feature dictionary generated from Chebyshev-spaced scalar parameters and quasi-uniform ridge directions. Train only the output linear map, or use the frozen layer as the first stage of a larger network, thereby eliminating hidden-layer backpropagation while retaining a constructive smooth-function approximation guarantee.
Replace a naively evaluated mixture of power-law experts with a Newton-envelope layer that computes all monomial magnitudes in log-space and subtracts their maximum before exponentiation. The layer exposes both a stabilized mixture value and soft dominance weights, allowing a downstream MLP to adapt to whichever scaling regime is active without overflow or hand-designed regime splits.
Replace random or k-means initialization of a k-expert router with a moment-based range finder on a calibration batch of hidden states. Estimate a low-dimensional second-moment subspace, enlarge it using one-free-index third-Hermite contractions, and fit the router's expert centroids and weights only in this resulting subspace. The router can then operate on projected hidden states while retaining an optional small residual adapter.
Replace a learned dense token-mixing matrix or residual-state transition with a sparse diffusive mixer whose Laplacian has a deliberately small largest Jordan block. Balance the two chain lengths around the central coupling/core, because the paper proves that this minimizes the worst defective transient among the tridiagonal family. Use a scalar residual step size to move the non-consensus spectrum inside the unit disk while preserving the sparse structure.
Use the network-space inverse inequality to choose derivative order, collocation resolution, and feature separation jointly instead of enforcing arbitrarily high-order residuals on an under-resolved network. This creates an anti-aliasing rule: a network whose parameters are separated by \(\underline h\) cannot represent high Sobolev frequencies without a factor \(\underline h^{-(r-s)}\), so derivative penalties above the resolvable order should be disabled or accompanied by refinement.
Represent periodic input-output behavior using a compact real vector of Fourier coefficients and learn an invertible neural map from input coefficients to output coefficients. Inference then obtains the input representation for a desired periodic output by a single inverse pass instead of iterative optimization through a nonlinear forward model, while the Fourier representation reduces sequence dimensionality when high-rate signals are spectrally sparse.
Add a periodic coarse optimization phase to SGD or Adam that operates on a compressed parameterization and returns a prolongated correction to the full network. Retain nonsmooth constraints or regularizers explicitly through a primal-dual update instead of relying on penalty smoothing. Accept the correction only when it improves a cheap fine-batch merit test, making the method useful even when the coarse objective is only approximately coherent with the fine objective.
Use enumerated weighing matrices as sparse orthogonal channel-mixing operators inside MLPs or residual blocks. Their ternary entries reduce multiplication to signed additions, while exact orthogonality prevents amplification or attenuation of feature norms; a trainable fixed-support version can recover expressivity without giving up computational sparsity.
Construct one reference Lanczos basis for a symmetric propagation operator H, then derive several seed-specific spectral responses for Q_a(H)x_0 through finite-band polynomial connectors. With degree-r seeds, each transformed basis vector uses at most 2r+1 neighboring reference basis vectors, avoiding a separate Lanczos factorization for every seed.
Use the slow-mode content of a neural network's local optimization dynamics to choose between a near restart and a deliberately larger restart concentrated in fast-curvature directions. The larger perturbation is predicted to recover faster when it has substantially smaller overlap with the slowest Hessian modes, producing an explicit Mpemba crossover in loss or validation recovery.
Replace a dense Gaussian or learned projection from dimension N to m with a normalized partial circulant projection generated by a single Gaussian vector. For K-sparse hidden states, the restricted-isometry guarantee predicts approximate norm preservation while reducing stored projection parameters from O(mN) to O(N). The projection can be evaluated with an FFT and should be combined with explicit top-k gating so that the sparse-input assumption is enforced.