Machine-learning ideas tagged Training dynamics in the ML taxonomy of the Math2NN corpus.
Augment gradient descent with stochastic relocations to uniformly sampled historical parameter vectors. In expectation, the optimizer receives a non-Markovian correction toward the running average of all previous iterates, which can suppress runaway directions and revisit earlier basins instead of remaining trapped in a sharp or unstable region.
Use the paper's Margulis-type law as a structural constraint for neural continuous-time dynamics: the number of isolated periodic latent trajectories with period at most T should grow like exp(hT)/T in a positive-entropy regime. This provides a falsifiable test for orbit collapse, excessive chaos, or spurious recurrence in neural ODE world models, rather than relying only on one-step prediction loss.
Estimate persistent frequencies in a neural-network training trajectory using a smooth weighted Birkhoff average instead of a rectangular moving average. Use the estimated frequency vector to detect low-order resonances between optimizer oscillations, gradient-noise cycles, and validation-loss oscillations, then trigger a learning-rate or momentum intervention before divergence.
Project each FFN residual update onto the tangent space of the current token residual direction before adding it to the stream. This preserves the component that changes representation direction while suppressing norm-only motion, which may reduce residual-norm drift and aggregation-induced representation collapse.
Equip a recurrent or state-space layer with multiple noncommuting transition operators and regularize the span of finite operator words applied to the input injection matrix. This discourages hidden directions that cannot be reached from the input and may improve long-range input influence, gradient propagation, and robustness under operator switching.
Construct an unrolled phase-retrieval network that begins with an isotropic Gaussian estimate rather than a spectral initializer. Retain the AMP residual correction and Onsager subtraction, but learn the scalar measurement denoisers and step sizes; use several random starts and select the iterate with the lowest measurement residual.
Replace a learned sequence-mixing matrix with a structured lower-triangular Sprugnoli operator whose square is exactly the identity. Applying the same operator in reverse reconstructs activations exactly, so it can be used as a reversible Transformer mixer or reversible channel permutation while retaining nontrivial long-range mixing.
Insert a constraint-aware observer between a neural state-space transition and its next prediction. The observer propagates latent event times, incorporates partial observations, and projects the result onto the set satisfying both lower-bound causality and upper-bound token-lifetime constraints, preventing impossible latent trajectories from entering the recurrent model.
Add a deliberate large-constant-learning-rate phase in which training loss is not forced monotonically toward interpolation. The phase is intended to calibrate shared, high-signal directions before the optimizer memorizes example-specific nuisance directions, and should be stopped when validation error is minimized even if training error remains high.
Build a neural feature-mixing block from a sparse exchange matrix and a fixed sequence of cluster-style mutations. Transpose-related channels share parameters through orbit tying, while positive multiplicative-additive updates are evaluated with log-sum-exp. The block is a concrete alternative to a dense MLP projection and can be tested at equal parameter count and compute.
Replace an unconstrained recurrent or residual transition with a block-structured contractive realization whose input-dependent multiplier is a direct sum of bounded branches. The resulting layer has a built-in non-expansive energy bound on the chosen normalized input domain, while still allowing different branches to respond to different coordinates or feature groups.
Store a finite library of successful robot configurations or action-conditioned waypoints and construct a smooth soft minimum of their distances. Use the negative distance gradient as a structured action prior, add a learned residual policy, and pass the combined action through a quadratic-program safety layer. This gives a neural controller an explicit attraction basin toward demonstrated solutions while preventing violations of known state constraints.
Attach a predictor from each deeper representation to the representation immediately below it, and penalize the Gaussian KL divergence between the predicted lower-layer state and the actual lower-layer state. Learn or estimate one positive variance per layer so easy, low-noise layers receive high precision while intrinsically uncertain layers are not forced to fit their targets exactly.
Attach a symbolic sparse head to a neural encoder instead of using a dense final MLP. The head evaluates a library of learnable power-law and interaction terms on nonnegative learned features, jointly optimizes linear coefficients and exponents, and removes inactive terms with coefficient sparsity. This should provide a compact model with better relative-error behavior on positive targets spanning several orders of magnitude.
Model stochastic training or recurrent inference as a random dynamical system and penalize the exponential growth of volumes transported by its Jacobian. This converts the paper's entropy and volume-growth relation into a computable regularizer that discourages chaotic sensitivity while retaining directions needed for fitting.
Replace or augment a mixture-of-experts router with a relative transverse-curvature score computed between experts, rather than relying only on the router MLP logits. Experts that provide a broader, less stiff local response in task-relevant directions receive higher routing probability, while common nuisance or spectator directions cancel from the comparison. The score is invariant under a common linear reparameterization of the routing coordinates and can be restricted to a low-dimensional…
Equip a latent world model with a learned positive-definite state-dependent metric and penalize violations of one-step contraction under the predicted dynamics. Use the paper's metric-geodesic energy as an auxiliary consistency loss between clean and perturbed latent rollouts, making the model more robust to observation noise and compounding prediction errors.
Replace an ordinary dense or floating-point eigendecomposition of small Hessian or Fisher blocks with a sequence of rational Jacobi rotations. The rotations preserve Euclidean norms and can be stored using fixed-point coefficients, while approximately diagonalizing curvature so the optimizer can use separate coordinate-wise step sizes. This is especially relevant to low-precision training and blocks with mixed-sign curvature.
Replace an expensive global resolvent calculation for a recurrent or state-space transition operator by measurements on overlapping finite patches. Penalize patches whose shifted operator has small minimum gain, while adding the paper's explicit O(1/n) truncation penalty so that increasing the patch size produces a predictable tightening of the stability certificate. This targets non-normal transient amplification that is invisible to ordinary eigenvalue or spectral-radius regularization.
Replace a single-step spectral-radius diagnostic in a recurrent network with a multiscale induced pressure computed from return trajectories. Separate return branches whose Jacobian products remain close to the limiting dynamics from transverse branches that create rapid growth in trajectory complexity, then reduce recurrent gain or optimizer step size when the transverse pressure exhibits the predicted square-root rise near a neutral bifurcation.
Run multiple neural-network parameter trajectories in parallel and define divergence, NaNs, loss explosions, or trust-region violations as absorbing failure events. Whenever one replica fails, replace it with a copy of a uniformly selected survivor while tracking the time since its last replacement. This creates an empirical quasi-stationary distribution of robust training states instead of relying on one potentially unstable trajectory.
Use three learned state-transition operators corresponding to three data axes, and train them to satisfy the paper's pullback-style interchange rule. For every local pair of axes, two successive updates should reach the same square state; for triples of axes, all six update orders should agree. This reduces sensitivity to scan direction and limits long-horizon drift caused by inconsistent local transitions.
Use a small population of neural parameter vectors and replace isotropic random perturbations with the paper's masked affine move toward the current best candidate. Select the mask probability and migration distance from the closed-form expected step-length and active-dimensionality formulas, allowing large exploratory moves early and progressively focused moves later.
Partition the network output into blocks according to their estimated local controllability order and replace the ordinary residual norm by the anisotropic gauge q_p(r) = max_i ||r_i||^(1/i). Train an inverse network or unrolled solver with blockwise target tolerances ||r_i|| approximately less than or equal to rho^i, so directions reachable only through higher-order changes are not incorrectly treated as equally first-order errors.