Machine-learning ideas tagged Regularization in the ML taxonomy of the Math2NN corpus.
Compute a translation- and rotation-robust perimeter feature from the Euler Characteristic Transform and append it to learned shape features. Unlike a finite-radius ECT comparison, the point-anchor subtraction cancels the constant Euler-characteristic tail exactly, eliminating the need to tune a spatial cutoff.
Correct minibatch or trajectory-based categorical entropy estimates using the paper's power-law occupancy asymptotic. The corrected estimate adds back entropy lost through unseen rare categories, with the correction magnitude inferred from the number of distinct observed categories and an estimated tail index.
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.
Train a recurrent or state-space neural model with an information regularizer that uses trajectory-dependent predictive information at low observation noise but switches toward instantaneous mutual information as sensor noise increases. The switch is driven by an online estimate of the relative reliability of transfer entropy and instantaneous dependence, rather than by a fixed hyperparameter. This should prevent noisy histories from forcing the latent state to memorize unreliable temporal…
Apply the paper's mechanism-contrast idea to ReLU decoders by requiring each piecewise-affine branch to produce a detectable and distinctive change across at least one activation boundary. Penalize branches with vanishing Jacobian jumps or nearly identical boundary signatures, discouraging observationally interchangeable decoder mechanisms.
Compress representations of graph fragments according to their empirical edge-connection rank instead of using a generic hidden dimension. For fragments with t open ends, learn only the quotient space of boundary behaviors that remain distinguishable after gluing, producing a compositional graph network whose boundary-state dimension is capped by an estimated R^t.
Parameterize a large linear layer as a sum of binary tensor products, W = Σ_l A_l ⊗ B_l, and regularize a factor-level upper bound on its top-k singular-value sums. The bound controls all Ky Fan norms of W while requiring SVDs only of the small factors, making it suitable for tensorized MLP or attention projections.
Use the paper's derivative-dispersion mechanism as a neural regularizer: the input-dependent forcing should produce different derivatives in different hidden directions. Penalize collapse of the Jacobian of the forcing map while retaining a contracting recurrent transition, so hidden states do not converge to a low-dimensional manifold caused by nearly parallel inputs.
Replace unconstrained output coordinates with a neural parameterization whose outputs are valid monotone profiles by construction, analogous to representing a Young diagram through nonnegative ordered row increments. Train the network against an explicit energy or negative log-probability while preserving the feasible geometry, rather than relying on penalties that permit invalid intermediate states.
Use attractor separation and noise-induced basin coalescence as a robustness test for recurrent networks with multiple learned memories or modes. Estimate the smallest perturbation amplitude at which initially distinct hidden-state attractors become geometrically indistinguishable, then train or operate below that threshold with a safety margin.
Replace a single recurrent transition with a finite bank of candidate positive linear transitions and use a minimax controller to choose the feedback action at every time step. The controller evaluates candidate successors, selects the action whose worst-case predicted cost is smallest, and clips the action to preserve nonnegative hidden states. This should make an SSM or RNN less sensitive to transition-matrix mismatch and long-horizon disturbances.
When the Schrödinger generator is learned, regularize its spectrum and eigenvectors so that the magnitude trajectory remains well-conditioned for recovering hidden complex states. Penalize small singular values of the squared-eigenvector matrix and near-colliding eigenvalue pair sums, preventing a learned dynamical layer from becoming spectrally invisible or phase-ambiguous.
Add a latent factor map that is approximately 1-Lipschitz and require it to preserve important scalar 1-Lipschitz observables of the data distribution. Approximate the universal quantifier with an adversarial bank of neural probes, rewarding the encoder for retaining distributionally stable information while discarding high-frequency or sample-specific detail.
Add a degree-conditioned neighborhood-profile penalty to a GNN so that its effective message-passing graph has a controlled hub-neighborhood trend. The regularizer can either target a rank-one null profile, where neighbor degree is approximately independent of root degree, or deliberately target a learned/reference logarithmic trend when preferential-attachment-like structure is useful.
Regularize a learned self-adjoint contraction so that its eigenvalues move toward 0 or 1 rather than accumulating in the transition interval. This suppresses ambiguous mixing modes and can enable a smaller binary spectral approximation at inference.
Compute one-dimensional persistent homology and minimum-norm harmonic representatives, then use their absolute edge coefficients as topology-aware saliency in a graph transformer. Add the saliency to attention logits or use it as a soft regularizer so the model preferentially propagates information along edges that are essential to persistent cycles.
Insert a differentiable layer that enforces multiple affine consistency constraints by running several short strings of relaxed projections and averaging their outputs. Change the strings and weights across training steps, but impose bounded string length, positive averaging weights, and an almost-cyclic coverage rule so every constraint is revisited regularly. This creates an architecture-level analogue of dynamic string-averaging rather than applying one fixed projection order.
For a learned recurrent or state-space model, estimate leading Koopman or transfer-operator modes and force their evaluations on a small set of latent states to be linearly independent. This transfers the paper's generic invertibility construction and discourages duplicated, weakly observable, or spectrally collapsed dynamical modes, potentially improving long-horizon prediction and interpretability.
Regularize a generator so that the Gram determinant of its Jacobian with respect to Gaussian latent noise rarely becomes very small. This should reduce latent-space collapse and make the generated distribution more regular, improving the chance that small Wasserstein or MMD errors correspond to small density-level errors rather than narrow singular spikes.
Augment a local convolutional block with a fractional nonlocal-gradient branch that aggregates directional feature differences over multiple spatial scales. The residual branch gives each location access to long-range variation while preserving the property that constant feature fields produce zero response. A learnable residual gate allows the network to suppress the branch if nonlocal interactions are unhelpful.
When clients optimize the same publicly known pair of losses but have private trade-offs, protect only the ratio of objective weights rather than the complete weight vector. Communicate a ratio-conditioned mixed gradient or controller statistic, with sensitivity defined over bounded ratio changes. This can reduce the required privacy noise when common rescaling of all objective weights carries no meaningful private information.
Learn a low-dimensional active frame for a neural scalar quantity on a curved latent manifold, rather than averaging gradients in unrelated ambient tangent spaces. Use the frame as the only input to a low-rank adapter or as a constraint on fine-tuning updates, with parallel transport making gradient statistics comparable across samples.
Train a learned latent transition not merely to fit one-step data, but to require only a small operator correction before its selected spectral modes become exact eigenmodes. The correction is a measurable backward error, so the regularizer penalizes models whose apparent eigenstructure is highly sensitive to noise or finite-sample error. At inference time, the correction norm can trigger conservative rollout or mode suppression.
Attach a small dynamical observer to a neural ODE, RNN, or state-space model and make it estimate only a task-relevant functional of the hidden state, such as logits, value features, or control-relevant projections. Use an incremental quadratic constraint and a bounded-real penalty to make the observer robust to hidden-state nonlinearities and input disturbances, instead of reconstructing the full latent state.