Machine-learning ideas tagged Inference speedup in the ML taxonomy of the Math2NN corpus.
Replace Cox-de Boor evaluation of each cubic B-spline edge activation with its fixed truncated-power expansion. Normalize each scalar edge input to a bounded knot coordinate, evaluate the five shifted cubic positive-part terms in parallel, and contract them with the learned spline coefficients inside one fused kernel.
Construct a residual network from independently attachable modules, but permit only a number of modules whose aggregate feedback gain lies inside a delay-dependent admissible interval. Estimate deployed end-to-end latency and each module's local Jacobian gain, then reject or bypass additional modules when the predicted delayed-loop stability boundary is crossed. This turns variable-width or depth scaling into a falsifiable control problem rather than an empirical choice.
Replace the standard unit-step modern Hopfield retrieval update with a relaxed step using theta greater than 1, while restricting theta to the theoretically safe interval (0,2). The relaxed map has the same fixed points as ordinary attention and provably decreases the Hopfield energy, so it can move farther toward an attractor per iteration without changing the retrieval objective.
Distill the expensive inner minimization over state-estimation errors into a neural correction term that predicts the robust barrier drift, then fine-tune the correction using differentiable closed-loop rollouts. This retains the robustness mechanism while reducing the repeated optimization cost and allowing less conservative behavior than fixed analytic uncertainty bounds.
Use the noisy diffusion state to estimate posterior probabilities over data clusters, then route the denoising computation to one or a few cluster-specific score experts. Each expert stores a low-rank covariance or low-rank score correction, so after posterior commitment the denoiser operates at the cluster intrinsic dimension rather than ambient dimension. During the early mixing phase, retain a shared full-dimensional score or evaluate a small weighted set of experts.
Replace fixed-grid stochastic quantization of each tensor block with an adaptive ordered codebook selected under both an entropy budget and a maximum number of codepoints. Within every interval between adjacent codepoints, use unbiased stochastic interpolation, so the quantized block remains unbiased while the emitted symbol distribution becomes easier for arithmetic or Huffman coding to compress. The representation should reduce actual compressed bytes at fixed MSE, or reduce MSE at fixed…
Attach a value-based stopping controller to any verifier-guided refinement loop. After each generated answer and verifier evaluation, estimate the value of accepting the current output and the value of continuing for one or more additional refinements; stop when the expected gain from continuation is no larger than its compute cost. The controller learns a score-dependent stopping boundary instead of using a fixed iteration count.
Learn a low-dimensional Koopman operator from successive states of an iterative neural system, such as debate agents, recurrent refinement blocks, or diffusion denoising trajectories. Use the magnitude of the subdominant eigenvalue to predict the remaining number of rounds required for disagreement to fall below a target tolerance, and stop computation when the predicted deadline is reached rather than using a fixed round budget.
Replace a full-resolution neural operator with an explicit multiscale reconstruction: a coarse predictor produces a low-resolution field, and a small corrector module combines its gradients and Hessians with learned rapidly varying basis functions. The model should recover fine detail without running the expensive backbone at fine resolution.
Use a second-order Runge-Kutta integrator satisfying the chain-tree condition b^T A c = 1/6 when the neural ODE output is an event threshold or separatrix crossing. The method remains only second order for general trajectories, but the paper predicts cancellation of the leading discretization bias in this nonlinear observable, potentially allowing larger inference steps at fixed threshold accuracy.
Replace a square dense projection in a Transformer or MLP with a trainable recursive butterfly matrix. The layer preserves multiscale channel interactions while constraining every complementary row-column block to rank at most k, reducing parameters and enabling recursive structured matrix-vector products. Unlike an arbitrary sparse layer, the construction has an explicit recursive factorization and a quasi-optimal approximation guarantee among matrices with the same butterfly rank.
Replace consecutive or randomly assigned transformed KV coefficients with groups whose variance-volume is approximately equal. Train one equal-size vector-quantizer codebook per group, so a fixed-width cache does not waste its low-rate budget by forcing high-variance and low-variance coordinates into badly mismatched groups. This is a drop-in quantization-layout change that can be applied to keys, values, or both.
Replace a dense neural-network weight tensor with a graph tensor network whose physical modes and internal edge ranks are specified by a sparse rank-adjacency matrix. Unlike tensor-train or hierarchical Tucker layers, the graph can contain selected cycles and skip connections between tensor modes, allowing the factorization topology to match correlations in the weight tensor. Fit the layer with GTN-SVD at a prescribed tolerance and compare accuracy, parameter count, and tensor-contraction…
Build a forward-mode second-order jet interpreter for the PINN and evaluate the entire PDE residual in one compiled graph. Each intermediate carries its value, first derivative, and Hessian with respect to the collocation coordinates, avoiding repeated nested reverse-mode autodiff calls for every residual component.
Replace dense self-attention by a sparse attention graph whose neighborhoods satisfy the paper's size-dependent expansion condition. This preserves a logarithmically controlled route for every token subset to communicate with the rest of the sequence, reducing quadratic attention cost without allowing disconnected or poorly mixed token groups.
Partition graph nodes into backward-equivalent classes and run message passing on the K-node quotient graph instead of the original N-node graph. If every node in a class receives the same aggregate message from every source class and shares the same local update map, class-constant node representations remain class-constant at every layer, making the quotient computation exactly equivalent to the full GNN on that invariant subspace.
Implement causal linear attention in chunks and combine chunk summaries with an associative scan instead of carrying the recurrent state through all chunks sequentially. This preserves the exact causal computation while reducing inter-chunk dependency depth from the number of chunks to its logarithm, enabling substantially more GPU parallelism for long-context training and prefill.
Train a neural trial function for an elliptic PDE using a very-weak residual in which all derivatives act on fixed smooth test functions rather than on the neural network. This eliminates second-order reverse-mode or forward-mode automatic differentiation and allows low-regularity activations while retaining a least-squares objective over many test functions.
Replace a dense directed attention matrix by a collection of K learned source-to-hub-to-target interactions. Each hub corresponds to a directed biclique, allowing many source tokens to communicate with many target tokens using O(NK) rather than O(N^2) pair interactions. The construction preserves asymmetric information flow and can be initialized from a graph cover of high-attention edges.
Replace an autoregressive rollout of a learned dynamical model with a branch-trunk factorization that predicts all future steps simultaneously. The branch network encodes the future action sequence, while the trunk network encodes the current state and query coordinates; their inner products produce the complete horizon. This removes repeated state updates during inference and gives a compact differentiable model for planning.
Attach a cheap risk score to each neural-network prediction and skip an expensive verifier, ensemble, diffusion refinement, retrieval call, or human review when the score is below a calibrated threshold. Independently audit a random subset of skipped examples using the expensive ground-truth procedure, and select the largest skip threshold whose exact confidence bound keeps the violation rate below a target budget.
Replace the standard CFG guidance multiplier in each DDIM step by a coefficient obtained from the terminal guided exponent. Given unconditional and conditional denoiser-derived states D_u and D_c, use r^{1+w}-r instead of w(r-1) on the guidance direction D_u-D_c; this preserves the same two denoiser calls and costs no additional NFE while suppressing low-noise residual blow-up at high guidance.
Compress each hidden layer by retaining directions that are simultaneously reachable from the observed input distribution and observable at the network output. Unlike PCA or SVD, the retained subspace is weighted by downstream task sensitivity, so high-variance but output-irrelevant directions can be removed while low-variance predictive directions are preserved.
Replace the final sequence of diffusion-sampler steps below a positive switching noise scale a with a single analytic normal-mode completion map. Run the existing solver only on [a, sigma_max], then use the denoiser at scale a to extrapolate to the requested terminal floor epsilon. This prevents the step count from growing like log(sigma_max/epsilon) and should preserve the base solver's order when a is coupled to the discretization size.