Insert channelwise normalization whose mean and variance are pooled over the full sequence, allowing a small-receptive-field convolutional labeler to access global sequence statistics without adding dilated convolutions or attention. Use this only for tasks where labels occur in long runs or depend on coarse global composition; retain per-position normalization for tasks requiring strict locality.
Use the paper's finite-width O(n^{-1/2}) Gaussian-process approximation bound as a width-budgeting rule rather than choosing every hidden dimension uniformly. Estimate an architecture-specific constant for each layer or attention contraction, then allocate width according to the smallest dimension satisfying its allowed distributional error. This should produce narrower models at comparable GP-like behavior, or permit the same parameter budget to be concentrated in the layers where finite-width…
Replace parameter-count or validation-loss-only model selection with a singular-complexity score based on the evidence scaling of each candidate neural network. Select or prune architectures using \(n\widehat L_n+\widehat\lambda\log n-(\widehat m-1)\log\log n\), which can prefer overparameterized but highly redundant networks when their effective singular complexity is lower.
Partition a neural state or feature vector into blocks and identify directed dependencies between blocks from one-step transition data. Use the inferred design structure matrix as a hard mask or soft gate on recurrent, state-space, graph, or mixture-of-experts couplings, replacing a dense unconstrained interaction matrix with a data-supported sparse graph.
Represent the quadratic OT potential with a strongly convex input-convex neural network and train it by matching the distribution of its gradient pushforward to the target distribution in a weak dual metric. Median-center the potential on every minibatch so that optimization does not waste capacity or suffer instability from the additive constant ambiguity. The paper's stability inequality predicts that this can produce a more stable potential estimate than directly optimizing a transport-map…
Generate many random symmetric decision trees and encode each input by the one-hot indicator of its reached leaf. Use the resulting fixed random feature vector as an additional input to an MLP, or train only a ridge/linear prediction head on it. The tree ensemble's Gaussian-process-limit interpretation predicts that increasing the number of independent trees should approximate a stable kernel while avoiding MCMC and difficult optimization over discrete split structures.
Treat the maximum dependency distance faithfully modeled by a finite neural architecture as an emergent correlation length, and estimate how it grows with depth, state size, or attention span. Fit the exponent \(\kappa\) and use it as an architecture-selection signal: a model with larger \(\kappa\) should acquire long-range competence more efficiently at equal parameter or FLOP budget.
Insert a magnitude-only bottleneck whose output is the absolute value of a random independent-feature expansion of the latent vector. Train a decoder to reconstruct the latent representation or input modulo one global sign, while explicitly rejecting feature distributions whose normalized L1 mass is too small. The module provides a controlled way to obtain sign-invariant representations without allowing arbitrary coordinate-wise sign loss.