Machine-learning ideas tagged Optimizer in the ML taxonomy of the Math2NN corpus.
Separate a neural network into nonlinear hidden parameters and a linear output layer. Solve the output layer exactly by least squares, then update hidden parameters with a truncated-pseudoinverse Gauss-Newton step that discards numerically singular directions.
Replace an unpreconditioned conjugate-gradient solve for a damped Gauss–Newton step with a two-level algebraic preconditioner derived from local Jacobian-row supports. Use overlapping local parameter blocks as Schwarz subdomains and a coarse basis containing low-energy local modes, so the optimizer can correct both localized and globally coupled parameter errors.
Replace abrupt optimizer preconditioner changes with a metric trajectory that moves the smallest affine-invariant distance needed to reach a target generalized Hessian condition number. During training, optimize a short horizon of log-diagonal or block-SPD metrics using a terminal curvature penalty and an intrinsic kinetic regularizer, then execute only the first metric in a receding-horizon controller. The method should reduce oscillations caused by rapidly changing second-moment estimates…
Track an exponentially discounted approximation to the current min-max saddle gap and use it to control the optimizer of a GAN or adversarial learner. If the recent gap rises, reduce both players' step sizes and clear stale momentum; if it falls consistently, cautiously increase the step sizes. Unlike ordinary loss EMAs, this signal measures whether each player is close to a recent best response and can detect equilibrium-tracking failure even when generator and discriminator losses look benign.
Use the paper's negative-semidefinite interaction curvature to detect and compensate for destructive coupling among layerwise learning-rate, momentum, or preconditioner mechanisms. Instead of independently tuning mechanism amplitudes, estimate their reduced curvature after hidden optimizer states relax, then apply a low-rank trust-region step or freeze mechanisms whose interaction curvature is too negative.
Use the diffusion graph's Dirichlet energy and almost-isometry inequalities to score whether a candidate minibatch preserves the low-frequency structure of losses, logits, or gradients over the dataset. Reject or augment batches that distort these quantities, producing a geometry-aware batch acceptance rule rather than relying only on random or loss-based sampling.
Replace the usual linear predictor in continuation of an implicit neural state with a fractional-power predictor fitted from recent states, then correct the prediction using a pseudo-arclength constraint. This is designed for equilibrium layers, implicit sequence models, or homotopy training schedules where the state Jacobian becomes nearly singular and ordinary Newton correction or fixed-point iteration becomes unstable.
Replace a fixed or hand-tuned learning-rate schedule with a slowly exponentially increasing schedule, and restart the schedule whenever the update norm grows at least as fast as the schedule itself. The restart preserves the current parameters but resets the learning-rate multiplier, allowing the optimizer to repeatedly approach the largest locally stable step size without requiring a Hessian spectrum or a reliable initial learning-rate guess.
Replace the usual unconstrained neural likelihood head with an unnormalized posterior potential that is linear in a learned coefficient vector over neural features. Optimize the exact partition-function-corrected posterior objective rather than only pointwise negative log-likelihood. This gives a globally convex final-layer problem and a positive-semidefinite covariance Hessian, reducing optimizer sensitivity and calibration failures.
Insert an active-set reduction step into a binary energy layer or Hopfield-style discrete optimizer. Coordinates whose signs are stable and whose local fields have a rigorous margin are frozen, while their interactions are folded into an induced bias and only the unresolved tail is updated. This preserves the exact conditional quadratic objective and can reduce dense interaction cost substantially when the state becomes polarized.
Choose gradient clipping thresholds from an explicit worst-case tail probability implied by an observed kurtosis bound, rather than using a fixed norm threshold or an empirical percentile. For a standardized centered gradient coordinate, the threshold achieving target outlier probability \(\delta\) is obtained by analytically inverting the paper's sharp tail formula.
Apply the paper's orthogonal-sketch idea to a trainable low-rank weight update or low-rank optimizer state. After every optimizer step, use a row sketch and Cholesky QR to recondition the factor basis, then compute the update with the standard tangent-space projector rather than an oblique sketched projector.
Replace repeated full-curvature factorizations when evaluating many Bayesian neural-network hyperparameter settings with one shared low-rank approximation of the data-induced precision update. Reuse the same curvature subspace while changing prior precision, observation-noise precision, or layerwise regularization, evaluating each candidate through a small rank-r matrix and Woodbury solves.
Constrain a neural parameter block to a bounded open domain and replace its Euclidean optimizer with a Riemannian gradient induced by the Hessian of the logarithmic barrier g=-log(-rho). The metric diverges near the boundary, so updates automatically become small when parameters approach saturation or an invalid region, while the logarithmic exhaustion has bounded intrinsic gradient.
Replace repeated multi-task training runs at different loss weights with pseudo-arclength continuation over stationary solutions of the weighted objective. Use homogeneous objective weights so that the algorithm can cross points where the conventional ratio of task weights diverges, then store the resulting network checkpoints as an approximate Pareto set.
Replace standard heavy-ball momentum with an update derived from a discrete kinetic-minus-loss action and a discrete viscous force. The force discretization produces a rational damping factor that remains controlled over a specified range of step sizes, potentially reducing oscillations and instability without Adam-style second-moment state.
Use the paper's non-permutation-invariant overshoot bound as a runtime guard for large learning rates. A proposed step is accepted only if its predicted overshoot contribution is compatible with the observed gradient residual; otherwise the optimizer clips or shrinks the step, preventing isolated very large updates from causing delayed divergence.
Use the observed power-law decay of a scalar training signal to estimate the effective fractional order of the optimization dynamics, instead of choosing the memory exponent by hand. Then run a fractional-memory optimizer with the estimated order, allowing the algorithm to use stronger long-range memory during slow plateaus and weaker memory when the loss relaxes rapidly.
Replace the ordinary gradient of a spatially indexed parameter tensor by a Fourier-domain inverse-metric gradient. FFT the gradient over its spatial dimensions, divide every frequency by a positive spectral symbol, inverse FFT, and then apply the optimizer step. Use a Bessel/Sobolev symbol as a parameter-free baseline and optionally estimate a task-specific symbol from gradient power spectra.
Treat a coupled neural training loop as a delayed feedback system with two hard delays and two first-order implementation filters. Estimate the dominant coupled Jacobian mode and use the characteristic equation to distinguish a recoverable delay-induced oscillation from a filter-induced instability; then reduce stale-gradient delay only in the former case, and slow or retune EMA or relaxation filters in the latter.
Track the implicit l2 regularization induced by adversarial SGD and explicitly correct it when the optimizer drifts toward an undesirable ridge strength. Apply the correction first to the final linear head or a low-dimensional adapter, where feature covariance and ridge estimates are tractable.
Replace fixed-strength projection or constraint-repair steps during low-rank neural fine-tuning with a regularized affine subproblem whose damping is proportional to the current distance from the model manifold. Use strong damping when a gradient update leaves the low-rank manifold substantially, then automatically remove the damping near a clean intersection so that the method can recover higher-order local convergence. This is suitable for LoRA-style updates, structured matrix compression…
Represent a parameter objective locally as a difference of convex terms, compute approximate proximal points for both terms, and update parameters using the difference of their high-order Moreau-envelope gradients rather than the raw DC gradient. Start with the quadratic case p=2, then test p=4 as a sharper penalty for large proximal residuals; solve each proximal subproblem with a small fixed number of inner steps and decrease the smoothing scale during training.
Wrap a learned residual policy or neural world-model controller around a stabilizing LQR feedback law, and permit sampling-based action refinement only when its estimated Monte Carlo and temperature errors fit inside a Lyapunov perturbation budget. Increase the rollout sample count, reduce temperature, or fall back to the baseline LQR action when the budget is violated. The controller should therefore trade computation for a measurable reduction in unstable or unsafe rollouts.