Machine-learning ideas tagged Optimizer in the ML taxonomy of the Math2NN corpus.
Calibrate the two blend coefficients directly from a context trajectory rather than using gradient descent. The one-step prediction problem is a two-variable ridge regression, making per-task adaptation nearly free and suitable for zero-shot or few-shot system identification.
Wrap a nominal gradient-based optimizer with a diagonal sign matrix that flips updates independently for parameter blocks, while a scheduler tests candidate sign configurations using short-horizon decrease of a Lyapunov-like training energy. The wrapper never changes the magnitude of the nominal update, and when the effective sign pattern is constant, it should recover the behavior of the correctly oriented nominal optimizer after a finite search period.
Assign separate sharpness or temperature parameters to two nonlinear subnetworks and anneal them according to a directional chart instead of driving both to their singular limits at the same rate. The optimizer explicitly tracks the ratio of the two scales and changes the schedule when the local Jacobian approaches a stability or bifurcation boundary. This tests whether the order and relative rate of sharpening, rather than only the final activation shape, controls optimization stability and…
Replace a fixed Adam update by an embedded Bogacki–Shampine RK3(2) proposal with a genuine accept/reject controller. Measure error between the two actual Adam parameter maps, rather than only between raw gradient estimates, and charge every gradient evaluation against the training compute budget.
Replace the single arbitrary autodiff derivative at a piecewise-smooth interface with a sampled conservative-field gradient envelope. For each minibatch and parameter point, collect gradients from locally reachable branches, average them as a convex combination, and use the resulting direction in a stochastic update. This is intended for architectures with routing, clipping, hard masks, or custom continuous branching where ordinary autodiff can select an unstable branch.
Treat each neural-network block as a local strength system and measure how perturbations in its input channels affect multiple output observables, rather than using a single gradient norm. Use the estimated maximum directional gain to cap residual updates or assign a layerwise learning-rate multiplier, preventing weak high-gain layers from destabilizing training while allowing strong layers to move faster.
Replace a fixed or heuristic noise-annealing schedule with one constrained by the FPU freeze-out scaling. In stochastic gradient Langevin dynamics, reduce the injected temperature slowly enough that residual parameter fluctuations remain below a target floor; if cooling is too fast, the optimizer should retain a measurable nonequilibrium variance analogous to the FPU residual energy.
Constrain the transition matrix of an RNN or linear state-space model to the paper's class Cρ instead of controlling only its spectral radius or spectral norm. The resulting transition has an explicit dilation certificate and satisfies ∥T^n∥ ≤ ρ for every time horizon, preventing exploding hidden states while retaining nonnormal dynamics that ordinary spectral normalization may remove.
Use the paper's norm-regularized conic dualization to impose PSD or SOS-style certificate constraints during neural-network training without forming Schur-complement or second-order-cone liftings. A neural dynamics model can be trained jointly with a polynomial Lyapunov or energy certificate, while the certificate subproblem is solved through accelerated updates in equality-constraint dual variables.
Add a regularizer to a recurrent or state-space transition that makes its expansion along a learned one-dimensional direction approximately constant across hidden states. A learned potential can absorb state-dependent terms, implementing the paper's cohomology mechanism rather than forcing the raw Jacobian to be constant.
Replace the memoryless parameter update with a discrete generalized Langevin update whose friction kernel is a positive mixture of decaying modes generated or scheduled by a Loewner driving process. Inject correlated gradient noise using the same kernel, implementing the paper's fluctuation-dissipation mechanism instead of choosing momentum and noise independently. The method is intended for noisy minibatch training, where controlled colored noise can preserve exploration while suppressing…
Treat consecutive optimizer updates as a discrete dynamical system and monitor the dominant local multiplier of the parameter-update map. When an estimated real multiplier approaches -1, apply damping or reduce the learning rate, because the paper's mechanism predicts the onset of an alternating period-2 orbit before ordinary divergence is visible.
Replace generic projected-gradient iterations for equality-plus-bilateral constraints with an active-set semismooth Newton layer. Each iteration fixes currently active lower and upper bounds and solves one structured saddle-point system, potentially converging in a few iterations when only a small subset of constraints is active.
Train a neural predictor using an explicit primal-dual gap instead of only a constraint residual. The gap measures objective suboptimality and constraint violation together, and can provide a principled per-example stopping rule for inner optimization or test-time refinement.
Replace Langevin or random-walk sampling for a strongly log-concave neural subproblem with randomized Hamiltonian trajectories. Each iteration draws a fresh Gaussian velocity, integrates position and velocity for a random triangular or exponential duration, and discards the terminal velocity before the next refresh. The target is a regularized posterior over a convex neural-network head, where the paper's accelerated dependence on the strong-convexity parameter is applicable.
When a chosen sparse support is geometrically incompatible with exact orthogonality, temporarily optimize on a nearby off-diagonally perturbed Stiefel constraint rather than forcing a singular Newton system. Anneal the perturbation to zero after the active support has stabilized, using the paper's O(||Delta||_F) KKT guarantee to control the residual of the original orthogonality-constrained problem.
Construct a Lanczos chain for the neural-network vector field or hidden-state evolution, separately within bins of approximately constant loss, energy, or activation norm. Use the resulting Krylov complexity and Lanczos-coefficient growth as an early-warning signal for unstable training or long-horizon hidden-state amplification, then reduce the learning rate or recurrent integration step only in the unstable shells.
Add a positive completely monotone memory branch to an optimizer or recurrent state update, but retain an explicitly calibrated instantaneous gradient or input branch. Estimate the memory branch's finite-horizon coercivity and prevent the system from entering regimes where memory suppresses high-frequency corrections and causes slow or unstable training.
Replace a fixed learning-rate schedule with a BB curvature step projected onto an adaptively estimated stable interval. Use the enlarged gradient-descent stability range, approximately below 2/L for an L-smooth objective, but verify every aggressive proposal with a sufficient-decrease test and fall back to a smaller step when the local curvature estimate is unreliable.
Train a square orthogonal neural mixer while maximizing its entrywise fourth-power concentration. When optimization reaches a non-permutation stationary configuration, explicitly test rank-two row or column rotations and take a rotation with positive exact second variation, using the paper's constructive saddle-escape mechanism.
Use nested parameter-confidence sets to control how far a neural optimizer may move when its local loss dynamics are uncertain. Estimate a local linear model of parameter or gradient evolution, propagate a homothetic tube for possible next iterates, and impose a trust-region radius that shrinks when the estimated contraction margin is insufficient. This gives a model-based alternative to heuristic gradient clipping and predicts a sharp learning-rate boundary tied to the largest uncertain…
Add a response-sensitive regularizer to networks whose outputs should react predictably to a control input, using the stationary Markov sensitivity equation as a certificate. Instead of only penalizing large neural gradients, the method attributes amplification to the generator resolvent and can distinguish amplification caused by a nearly slow latent mode from amplification caused by uncontrolled parameter growth.
Use the feedbacked control-to-state norm as a conditioning diagnostic to adapt the optimizer step applied to recurrent residual outputs. When the estimated horizon amplification is large, reduce or precondition the residual-control update; when feedback makes it small, permit larger updates.
Model locally competing neural-network parameter basins as low-energy states with different effective multiplicities, and inject calibrated parameter noise to measure when the optimizer begins switching between them. Use the resulting pseudo-transition peak as a principled trigger for changing learning rate, noise, or regularization rather than relying on a fixed epoch schedule.