Neural Network Optimization

Overview

Neural network optimization asks how to adjust a model’s parameters so it learns useful patterns efficiently and reliably. Common approaches include stochastic gradient descent and Adam, matrix-aware and spectral optimizers such as Muon, curvature-based methods, learning-rate schedules, and techniques that reduce memory or computation. Researchers also analyze loss landscapes, scaling laws, and initialization, and explore alternatives such as local or forward-only learning and analytic solutions for restricted models. A central tension is that gains depend on architecture, data, and training budget, while faster loss reduction does not necessarily improve generalization; connecting practical performance to reliable theory remains an open challenge.

Latest papers

355 papers29 papers in the last four weeks, down 28% on the four weeks before. 0.2% of all new papers.

Nov 15, 2026

Barron Approximation and locally optimal weight densities for shallow neural networks

Mean field theories provide optimization results for shallow neural networks by analyzing the weight distribution in infinite width limits. Corresponding results for finite sized networks are obtained by particle approximations, for which sharp quantitative bounds are still an open problem. In this paper, we consider a modified mean field loss, which allows a more fine grained control over finite sized networks. We prove convexity and equidistribution properties, which directly lead to Barron type approximation results for networks sampled from local loss minimizers. We demonstrate that particle approximations of the new loss function naturally lead to gradient descent methods with dropout type regularization.
Nov 14, 2026

NAP-Wave: A Neural Adaptive Plane-Wave Architecture for High-Frequency Helmholtz Problems

Training neural networks to simulate highly oscillatory wavefields presents optimization challenges, often hindered by spectral bias and poorly conditioned physics-based loss landscapes. To address these challenges and enhance convergence, we propose NAP-Wave, a joint neural architecture and training strategy that expands the target scattered wavefield in an adaptive neural plane-wave basis. The basis functions are extracted from all hidden layers and modulated by trainable plane-wave carriers, embedding the high-frequency oscillatory physics directly into the network's trial space. NAP-Wave is trained on a physics-informed loss built from the Lippmann-Schwinger integral equation, so no differentiation of the network with respect to its input is required. We develop a decoupled optimization strategy in which Adam updates the parameters of the neural plane-wave basis, while the linear combination coefficients are computed exactly by a least-squares (LS) solver. We couple this with an adaptive-depth training strategy in which the network starts shallow and grows progressively layer-by-layer. At each growth step, we avoid the loss spikes typical of dynamic network growth using the LS-computed coefficients, the warm-started parameters, and the preservation of the first- and second-order optimizer moments of the previous stage. Ablation studies confirm that the plane-wave basis, the exact coefficient solve, and the progressive schedule independently accelerate convergence and improve accuracy. Validated on exact-solution benchmarks, NAP-Wave achieves relative L2 errors of order 1%. On highly heterogeneous geophysical benchmarks, it achieves relative errors of 5%, 9%, and 10%, significantly outperforming published PINN methods in both accuracy and convergence time. To ensure reproducibility, the code is available at https://github.com/Hamza-Kamil/NAP-WAVE.
Nov 14, 2026

Grokking Can Be Steered: Selecting Desirable Solutions By Choice Of Learning Rule

During grokking, a network transitions from memorization to generalization, yet what shapes the neural implementation (circuit) of the generalizing solution and specifically the number of participating neurons (its size) remains unclear. We find that the learning rule, the combination of credit assignment algorithm and optimizer, controls this: On modular addition, every learning rule we study that groks recovers the same Fourier-based algorithm, yet implemented by circuits that differ in size. Why? A neuron stays in use when its credit is large relative to the non-gradient term of its own update, and this single condition unifies three routes to unused neurons: the coupled weight decay, Adam's εε and the decay of a trainable per-neuron gain. Equalizing the magnitude of credit assigned to each neuron, or scaling each neuron's weight decay by its credit, expands the effective circuit size, measured from the outgoing weights, from 60% to 99% of the layer's neurons at similar training and test accuracy. We vary which neurons share one equalization factor, from the whole layer to individual neurons, and find that the circuit stays concentrated for groups of eight or more and fully distributes only when every neuron has its own factor. Credit magnitudes are usually set by the normalization of standard optimizers: Nero, which normalizes per neuron, distributes the computation, whereas Adam and seven others, including RMSProp, Muon, and Shampoo, use other normalization schemes and retain small, concentrated circuits. Distributed circuits tolerate neuron deletion better, while concentrated circuits can be pruned further and transfer about twice as fast to a task that recombines the same features, largely by recruiting their unused neurons. Learning rule choice thus steers circuit size, trading robustness against compressibility and transfer speed without changing the learned algorithm.
Nov 14, 2026

Neural Dual Bounds: Valid-by-Construction JGLP Warm-Starts for MAP and Constrained MAP

We present Neural Dual Bounds, a calibration-free framework that produces deterministic upper bounds for MAP inference in probabilistic graphical models using a single neural-network forward pass. The framework trains a network to output the edge messages of a mini-bucket style upper bound on a given join-graph decomposition in a self-supervised manner by directly minimizing a smooth dual objective. The output of this network can be used at test time to warm-start a standard JGLP optimizer. Across diverse benchmarks, this warm start yields dual gaps that are orders of magnitude tighter than those from zero initialization, with the largest gains in the early-iteration regime. We also apply our framework for the constrained MAP (CMPE) task, where neural warm starts with test-time adaptation are comparable to the adaptive baseline on Bayesian networks and achieve the tightest dual gaps, by one to two orders of magnitude, on the grid and segmentation families.
Nov 10, 2026

CAFIN: CUDA-accelerated Adaptive Fisher Information Matrix Network Optimization via GPU-Native Kernel Fusion

Second-order optimizers improve deep neural network training using curvature information, but their practical efficiency is limited by GPU execution overhead. AdaFisher reduces Fisher Information Matrix (FIM) approximation costs through diagonal Kronecker structures and exponential moving averages, yet its PyTorch implementation requires sequential CUDA kernel launches, intermediate tensor allocations, and repeated memory accesses. We introduce CAFIN, a GPU-native CUDA kernel fusion framework that improves AdaFisher execution efficiency without changing its optimization formulation. The proposed approach progressively integrates FIM preconditioner construction, exponential moving average updates, gradient preconditioning, momentum accumulation, weight decay, and parameter updates into fused CUDA kernels. Shared-memory reuse, register-resident computation, and in-place operations further reduce memory traffic and intermediate allocations. Experiments using ResNet-18 on CIFAR-10, CIFAR-100, and Tiny ImageNet across NVIDIA GTX 1660 Ti and RTX 3060 Laptop GPUs demonstrate up to 78.4% lower isolated FIM construction latency and 2.1-26.9% lower optimizer-step latency. The implementation also eliminates approximately 18.91 MB of intermediate GPU memory allocations per optimizer step. Numerical validation confirms consistency within the predefined FP32 tolerance, while convergence accuracy remains comparable across evaluated datasets and random seeds. These findings demonstrate that CUDA kernel fusion can substantially accelerate second-order optimization while preserving AdaFisher's underlying mathematical formulation and learning behavior.
Nov 9, 2026

Spectral Unlearning: Transformer Structure-Preserving Updates for Language Models

Machine unlearning in large language models (LLMs) aims to remove specific learned knowledge, capabilities, or behaviors while preserving model utility. Existing methods have mostly focused on the design of unlearning objectives, while largely overlooking the intrinsic matrix structure of transformer weights. In particular, standard optimizers operate on independent scalar coordinates, thereby ignoring this matrix structure. We propose spectral unlearning, an update framework that combines standard unlearning objectives with matrix-structured optimization. We show that the spectral update corresponds to the steepest-descent direction for transformers composed of linear layers with normalized inputs and outputs. Our approach is applied on top of existing unlearning methods and evaluated across complementary axes of memorization, privacy, and utility, summarized via their harmonic mean. Spectral unlearning consistently outperforms a broad range of baselines and closes most of the gap to a retain-trained reference. Code is available at https://github.com/unlearn-ml/spectral-unlearning/.
Nov 5, 2026

On Lipschitz Explosion in Deep Neural Networks with Normalization: Consequences for Optimization and Adversarial Robustness

Can a network look stable at every layer and still be exponentially unstable as a whole? The Lipschitz constant of a neural network is a basic stability quantity appearing in optimization guarantees, generalization bounds, and robustness certificates. Deep networks naturally admit two such notions, input Lipschitzness, measuring sensitivity to data perturbations, and parameter Lipschitzness, measuring sensitivity to weight perturbations. We prove that, for deep networks with normalization layers, with batch normalization as the canonical example, both Lipschitz constants can grow exponentially with depth and hence with parameter dimension, even under strong per-layer norm control such as ∥Wℓ∥2≤1\|W_\ell\|_2\le 1, and already for linear activations. Thus, a mechanism widely used to stabilize training can create severe worst-case instabilities that are invisible locally, i.e., from layerwise norm bounds alone. Parameter-Lipschitz explosion turns Lipschitz-dependent nonsmooth optimization guarantees into exponential-in-dimension bounds for deep normalized networks, while input-Lipschitz explosion makes worst-case Lipschitz-based generalization and robustness certificates vacuous. The latter also yields a concrete rank-separation mechanism for adversarial vulnerability. The theory predicts that perturbations introducing new singular directions should be amplified much more strongly than perturbations at the same noise level that remain within the input's original singular subspace. Experiments on MNIST, Fashion-MNIST, and CIFAR-10 support this prediction, showing that rank-creating perturbations cause substantially sharper drops in accuracy, confidence, and margins than same-subspace perturbations.
Nov 5, 2026

An Educated Guess: Deriving Statistically Aligned Gradient Estimates for Zeroth-Order Optimisation

As novel hardware architectures are explored for the purposes of machine learning, including neuromorphic chips, asynchronous systems, and ASICs, effective learning algorithms are highly desirable. While backpropagation is the standard for differentiation on standard hardware, its requirement for global synchrony and exact differentiability limits the exploration of exotic algorithms on non-traditional substrates and devices. Zeroth-order (ZO) optimisation methods, such as SPSA and weight perturbation, offer a compelling alternative by estimating gradients through inference-only passes; however, these methods have historically suffered due to the `curse of dimensionality', where random perturbations become increasingly misaligned with the true gradient in high-dimensional spaces. In this work, we move beyond blind random noise by deriving zeroth-order search directions from activation geometry available during the forward pass. This leads to a family of Spectral Zeroth-Order methods, SZO-κκ, whose directions solve local response-based gradient-guessing objectives under explicit stochastic assumptions. These methods explore feature-supported directions that are more likely to produce informative loss responses. Using multilayer and convolutional networks, we analyse gradient alignment, bias-variance structure, and training stability of these methods relative to standard perturbation baselines and backpropagation.
Nov 5, 2026

Module-Aware Optimization for Graph Neural Networks

Different optimizers impose different geometries on parameter updates. Adaptive methods such as Adam and AdamW use coordinate-wise updates, while matrix-aware optimizers such as Muon utilize weight-matrix structure through orthogonalized updates. For Graph Neural Networks (GNNs), however, the effective geometry of its weight matrices is shaped not only by its learned channel-mixing weights, but also by the graph-propagated representations it acts on. Thus, even if an update is well normalized as a matrix, its effect on the layer output can be amplified when applied to graph-propagated features. To address this gap, we introduce GraphMuon, a graph-aware optimizer that adapts Muon's matrix updates to the covariance structure of graph-propagated layer inputs. GraphMuon performs orthogonalization in a metric defined by these propagated representations and normalizes step sizes according to the update's induced effect on the layer representation. We show that theoretically, vanilla Muon can suffer from graph-induced update amplification, while our GraphMuon can mitigate this effect. As such, our GraphMuon yields a graph-metric analogue of Muon's polar update step. Empirically, GraphMuon achieves 1.32×1.32\times-2.42×2.42\times validation-loss time-to-target speedups over AdamW on three large-scale OGB benchmarks. Across 15 dataset-architecture configurations spanning node classification, inductive graph classification, and graph regression, it achieves lower time-to-target in 13 cases, while maintaining competitive or superior predictive performance. Code is available at https://github.com/guy-hadad/GraphMuon.
Nov 5, 2026

Projection-Free Gradient Tracking Neurodynamic for Distributed Nonnegative Sparse Optimization

This paper proposes a continuous-time, projection-free neurodynamic system based on gradient tracking for distributed nonnegative sparse optimization. Unlike traditional methods, our approach avoids projecting states onto the feasible set. Instead, it employs gradient tracking to estimate average gradients under heterogeneous costs, coupling primal and dual dynamics such that feasibility and sparsity emerge naturally from the equilibrium conditions. The system ensures unique solutions through a locally Lipschitz vector field and guarantees that every equilibrium corresponds to an exact KKT point. We establish rigorous convergence guarantees for convex objectives through three distinct analytical layers. These include a centralized kernel rate, a computable distributed rate derived from Jacobian analysis without restrictive stability assumptions, and a global exponential stability proof based on Lur'e system theory. Our analysis quantitatively separates these layers, revealing that the convergence rate is primarily limited by graph sparsity and remains robust to variations in momentum time scales. Numerical experiments on strongly convex quadratic problems validate the theoretical rates, demonstrate effective active set recovery, and show that the system maintains stable tail convergence rates even under strong local heterogeneity.
Nov 5, 2026

A Bilevel Neurodynamic Approach for Distributed SCAD-Regularized Sparse Least Squares

This paper proposes a bilevel particle swarm optimization (PSO)-neurodynamic method for solving the distributed sparse least-squares problem regularized by the smoothly clipped absolute deviation (SCAD) penalty to global optimality. The SCAD penalty is nonsmooth and nonconvex, so existing distributed neurodynamic methods have not been extended to it directly. Through two-time-scale singular perturbation analysis, a rigorous full-trajectory convergence theory with explicit convergence rates is established for the inner collaborative multi-agent neurodynamic system without requiring the second-order sufficient condition. A theory-guided inner-outer collaborative optimization framework is further developed, in which a red-light green-light sensitivity mechanism is embedded in the outer PSO search to enhance global search while concentrating computational effort on the SCAD shape parameters, thereby improving recovery accuracy and computational efficiency. Simulation results on four sparse-recovery scenarios demonstrate that the proposed method accurately recovers the true sparsity pattern and achieves faster convergence and lower relative error than competing neurodynamic methods.
Nov 3, 2026

Positional Encoding is All You Need for Scalable Equivariance Constraint Relaxation

Equivariant neural networks incorporate known symmetries of the data, such as permutations, rotations, or translations, directly into their architecture, often improving sample efficiency and generalization. However, strict equivariance can make optimization difficult, for example, when the data only approximately satisfy the assumed symmetry, or when the equivariance constraints induce a challenging loss landscape. Recent studies have shown that relaxing equivariance during training by introducing non-equivariant dense linear layers can ease optimization in such cases and improve the performance of the resulting strictly equivariant model, after removing the relaxation layers at test time. While effective, this approach incurs a significant computational overhead, limiting its applicability in scalable architectures. In this work, we propose a more efficient constraint-relaxation method based on positional encoding (PE). Our framework injects a positional encoding with learnable scale that breaks the model's original symmetry. This positional signal is gradually annealed to zero, recovering the original equivariant model by the end of training. Empirically, we demonstrate significant improvements for translational and rotational equivariant models with only marginal compute and memory overhead.
Nov 3, 2026

Non-Interfering Asymmetric AdamW Updates for Constrained Exploration of Equifinality in Hybrid Models

Hybrid models let neural networks supply the unresolved parameters of a process-based simulator, and are fitted to observations that constrain aggregate fluxes and storages but not the internal partitioning that produces them. Such a fit is equifinal: configurations indistinguishable in the observed aggregates can disagree substantially about the unobserved internals, and therefore about the mechanism they attribute to the same data. We study evapotranspiration, whose total flux is observed while its split into transpiration and evaporation is not. A model ensemble built from random seeds alone spans only a narrow band of the partitionings the data admit. We produce the spread deliberately, training members under auxiliary repulsion on transpiration, subject to the fit to observations not degrading. Non-interference is usually arranged by making the auxiliary gradient orthogonal to the main one, but under AdamW the applied step is a preconditioned momentum direction, so orthogonality between raw gradients no longer implies it. We introduce Non-interfering Asymmetric Network Optimization (NANO), an update whose first-order effect on the main loss is exactly that of the main objective alone, whose auxiliary step is bounded, and which concentrates repulsion in the coordinates the main gradient leaves idle.
Nov 3, 2026

CAPSkip: Cancellation-Aware Prefix Skipping for Selective Backpropagation

Per-step selective back-propagation forwards every example of a mini-batch but back-propagates only some of them. Most rules rank examples by loss or gradient magnitude, but such scalar scores ignore cancellation between per-example gradients: the gradient removed by a set can be much smaller than its individual magnitudes suggest. We introduce CAPSkip (Cancellation-Aware Prefix Skipping). Among the examples the model already classifies correctly, CAPSkip skips the longest prefix, in increasing order of last-layer gradient norm, whose vector sum stays within a relative budget bb of the batch's last-layer gradient. In exact arithmetic this bounds the change of the raw final-layer gradient, using rank-one contributions already available from the forward pass; the bound does not extend to deeper layers, optimizer updates, or the training trajectory. Because the cumulative distortion is not monotone in the prefix length, CAPSkip scans all prefixes; a factorised Gram matrix makes the scan cheap and keeps it on the device. With activation reuse and selected-row backward, CAPSkip fine-tunes DeiT-S on CIFAR-100 for 103 epochs in 0.576x the time of 103-epoch dense training, and is +0.25 pp [+0.08, +0.42] above an approximately time-matched 59-epoch dense run (n=3).
Nov 3, 2026

Theoretical Analysis of Muon on Transformers for Next-Token Prediction

Muon has recently shown strong empirical performance in training Transformer-based large language models (LLMs), yet its optimization dynamics remain poorly understood from a theoretical perspective. In this paper, we analyze the training dynamics of Muon for next-token prediction (NTP) in Transformers. The training procedure consists of two phases: the preprocessing phase learns the output-value component, while the main phase learns the key-query component. For the preprocessing phase, we characterize the optimization trajectory exactly and show that the normalized spectral margin is optimal after each update. For the main phase, we establish nonasymptotic convergence rates for the training loss and the normalized margin gap under softmax attention. These results suggest that Muon's update rule can help models efficiently learn which tokens matter for prediction, offering a possible explanation for its strong performance in large language model training. The learned weight structure also yields a generalization guarantee for unseen contexts. Furthermore, we validate our theoretical insights with experiments on next-token prediction tasks.
Nov 2, 2026

Variational Functional Neural Networks (vfNN)

We present a variational functional neural network (vfNN) algorithm that optimizes physically motivated continuum functionals in integral form (e.g., Lagrangian, Hamiltonian, or free energy) to find the states of a system. That way, vfNN bypasses solving the equations of motion (a.k.a. the Euler-Lagrange equations), which are in the form of differential equations. To this end, our approach is fundamentally different from conventional physics-informed neural networks (PINN) that solve differential equations by minimizing a phenomenological loss function involving higher order derivatives and ad hoc weight factors. In contrast, vfNN merely evaluates lower order derivatives of the generating functional and replaces the weight factors with rigorous physical scales. This also fundamentally changes the loss landscape, and the desired solutions may become unstable. For this emergent nonconvex optimization problem, we develop a quasi-Newton optimization algorithm with simulated annealing that converges to the solutions of the corresponding Euler-Lagrange equations of the functional regardless of their stability properties. Our simulations using vfNN give accurate results for benchmark systems governed by linear or nonlinear ordinary differential equations.
Oct 30, 2026

Gradient Descent on Two ReLU Neurons: Global Landscape and Bifurcation Dynamics

Understanding how gradient descent learns features remains a central challenge in neural network theory. We study this question in a minimal multi-index setting that already exhibits rich multi-phase gradient dynamics: a well-specified two-neuron ReLU teacher-student model under isotropic Gaussian inputs, trained by population gradient flow from small initialization. We show that the dynamics are organized by two directions: the "easy" bisector direction, which carries the leading signal, and the "hard" splitting direction, which governs specialization. To characterize the loss structure and learning behavior, we analyze the population squared-loss landscape and show that every critical point with both student weights nonzero is either a global minimum or a saddle. We then track the gradient-flow trajectory, showing that student neurons first collapse toward a bisector saddle before escaping and specializing to teacher neurons. Our results provide both a landscape and dynamics account of the multi-phase symmetry-breaking behavior that arises in a simple multi-index model and under standard algorithmic and architectural choices.
Oct 28, 2026

Optimization-based lane-changing trajectory planning in mixed traffic with embedded input convex neural networks

Lane-changing (LC) maneuvers are highly challenging vehicle interactions, with substantial implications for traffic safety and efficiency. Developing effective LC strategies for connected and automated vehicles in mixed traffic is complicated by the uncertainty associated with surrounding-vehicle behavior and the computational burden of embedding realistic prediction models within optimization schemes. To capture these complex interactions while retaining tractability, this study proposes a data-driven optimization framework that embeds neural network (NN) surrogates for safety and lateral comfort within a Bézier-curve-based trajectory planner. Specifically, the safety cost during the LC process is quantified via a surrogate composed of an input convex recurrent neural network (ICRNN) encoding the planned trajectory, a fully input convex neural network (FICNN) embedding the LC duration, and a partially input convex neural network (PICNN) decoder fusing them with historical traffic states; separately, another FICNN surrogate approximates the optimal lateral comfort cost as a function of the LC duration. Both surrogates are embedded via linear programming (LP) formulations, avoiding the binary variables required by conventional mixed-integer programming (MIP) approaches for NN-embedded optimization. We further theoretically establish their exactness at optimality and extend the framework to problems with additional analytical objective terms and multiple surrogates. Experiments on a hybrid dataset combining real-world and simulated trajectories demonstrate that the proposed method achieves surrogate accuracy competitive with unconstrained baselines while substantially improving optimization efficiency over conventional MIPs. Closed-loop simulations further show that the proposed planners markedly reduce the weighted planning cost and consistently outperform baseline methods.
Oct 27, 2026

LRMoE: Demand-Side Load Re-Calibration for Efficient MoE Training on Memory-Constrained Nodes

Sparse Mixture-of-Experts (MoE) models offer scalable capacity for Web applications, but uneven routing creates computation and communication bottlenecks. Existing placement and replication methods redistribute this workload, yet additional expert copies compete with token-related buffers for limited GPU memory. We present LRMoE, a memory-aware MoE training system that jointly recalibrates backward demand and balances the retained workload. During forward execution, it plans retention quotas and expert placement under resource constraints while preserving all routed computations. During backpropagation, a load recalibration metric estimates each token-expert pair's contribution to the expert update and selects pairs within these quotas. A capacity-aware routing algorithm distributes retained pairs among eligible instances of their original experts. Physical compaction then forms smaller communication payloads and variable-size grouped GEMMs, reducing backward traffic, computation, and execution-buffer demand. Experiments on Qwen3-30B-A3B and GLM4.5-106B-A12B show that LRMoE reduces cumulative MoE-layer training time to fixed-loss targets by 21.2-44.8% relative to SYMI.
Oct 27, 2026

Nonlinear Direct Feedback Alignment for Scalable Backpropagation-Free Training

Backpropagation (BP) remains the predominant algorithm for training deep neural networks, but it is biologically implausible and suffers from the backward locking problem, constraining parallelization across layers. Direct Feedback Alignment~(DFA), which propagates the global output error directly to each hidden layer via fixed feedback matrices, has emerged as a promising alternative for addressing these limitations. However, a significant performance gap persists between DFA and BP, which we attribute to the limited ability of fixed feedback matrices to capture the nonlinear transformation from hidden representations to the network output. To address this limitation, we propose Nonlinear Direct Feedback Alignment (NDFA), which constructs each hidden-layer feedback using a lightweight nonlinear surrogate subnetwork. To enhance the effectiveness of the feedback throughout training, we further propose a learning strategy to update the surrogate subnetwork, thereby better approximating the underlying nonlinear transformation. Extensive experiments across residual networks and vision transformers on CIFAR-10, CIFAR-100, SVHN, and ImageNet datasets demonstrate that NDFA consistently outperforms existing DFA-based methods and substantially narrows the performance gap to BP, while maintaining efficient parallel training. The code is available at https://github.com/YanFan321/NDFA.
Oct 26, 2026

The proton structure one mode at a time: a Neural Tangent Kernel analysis of a DIS fit

We study the training dynamics of a feed-forward neural network (NN) through the analysis of the neural tangent kernel (NTK) in a phenomenologically relevant fit of parton distribution functions (PDFs) from deep inelastic scattering (DIS) data. The exact integration of the gradient flow derived in a previous single-flavour analysis is generalised to the multi-flavour case, yielding an analytical perspective on the relationship between data and methodology during training. The Shapley value recently used in the context of PDFs is derived within our statistical framework for observables that depend linearly on the PDFs, and we outline a geometric interpretation of training. We expose the role of gradient-based optimisation algorithms and of the parametrisation in the regularisation of the ill-posed inverse problem that affects the determination of PDFs from experimental data, demonstrating that the regularisation amounts to the incremental inclusion of NTK modes, each learned at the timescale set by the corresponding eigenvalue. We find that only a few leading modes of the NTK are necessary to reconstruct the target PDFs, while the noise of the data is carried by the sub-leading ones learned at later stages of training. Finally, we consider the impact of the initialisation on the training dynamics, and recover a transient behaviour observed in the previous analysis and in the machine learning (ML) literature. Beyond sharpening our understanding of the training dynamics, our framework can be applied to other parametrisations (such as fixed functional forms) and ML applications in physics, as well as to stress tests in the presence of inconsistent data.
Oct 24, 2026

Reactive Optimizers

We study a reactive optimizer that uses a pretrained language-model agent to inspect training dynamics and rewrite executable optimizer code during a single training run. The agent can change update rules and schedules for individual parameter groups, without optimizer-specific meta-training or memory transferred from other runs. We evaluate this approach, OPAL, on language modeling and image generation when a recipe has not been tuned for the target setting. Across the evaluated settings, it matches or improves on selected fixed recipes. Analysis of its decisions shows changes to learning-rate schedules, optimizer structure, numerical precision, and momentum. Controlled experiments also show adaptation to repeated data corruption and reduced overfitting. We report these behaviors together with the agent's computational cost and implementation requirements.
Oct 24, 2026

Optimizing Physical Neural Networks through Device Variable Selection

Physical neural networks (PNNs) leverage device physics in electronics, optics, and related domains to enable energy-efficient neural networks. Broken-isomorphism PNNs replace linear operations with multivariate physical transformations, modifying how input signals are mapped to device variables. We exploit network architectures emerging from these mappings by treating the routing between signals and input variables of physical devices as trainable parameters. This approach expands device-level tunability of the physical transformations while, at the system level, demonstrating improved expressivity of PNNs with negligible additional inference overhead. This discrete routing is optimized by training a categorical distribution using Gumbel-Softmax reparametrization. We evaluate this approach using experimentally calibrated differentiable models of multivariate electronic RNPUs for MNIST and CIFAR-10 classification. Using variable selection, we improve the MNIST classification accuracy for a small CNN model from 94.5% to 96.9% and the accuracy of CIFAR-10 from 62.8% to 68.0%. Our method uniquely leverages the induced biases of broken-isomorphism PNN architectures, paving the way towards more efficient physics-based neural network inference.
Oct 22, 2026

Training Mixed-Monotone Neural Networks for Verification Without Set Propagation

Most formal verification approaches for neural networks conservatively enclose the output set of the neural network. However, computing a tight output enclosure is in general computationally hard. We sidestep this challenge with a novel neural network architecture that realizes mixed-monotone functions as the difference of monotone subnetworks. By utilizing the theory on mixed-monotone systems, this reduces the computation of a tight output enclosure from complicated set propagation to two simple forward passes of opposite corners of an input interval. Moreover, because the size of the resulting enclosure is a closed-form function of the weights of the neural network, we can directly minimize it during training. We prove a lower bound on the size of any such enclosure, given by the total variation of the represented function generalized to several inputs, which makes it possible to compare the tightness of the enclosure against the optimum. Our evaluation on regression and dynamical-system benchmarks demonstrates that our training reduces the enclosure size to within 0.5% of the lower bound without sacrificing prediction accuracy. Computing our enclosure is orders of magnitude faster than CROWN and αα-CROWN and on deep networks it is also tighter than zonotope propagation and CROWN, while its size exceeds that of αα-CROWN by only 5.1%.
Oct 21, 2026

Forward-Only Learning for Shared-Kernel Wavelet Neural Networks without Backpropagation

Backpropagation remains the dominant training paradigm, using backward gradients and local updates that can suffer from vanishing or exploding gradients in deep networks. Although the Forward-Forward algorithm avoids conventional backpropagation, it remains mainly local or layer-wise, motivating global whole-network learning using forward evaluations only. For shared-kernel wavelet neural networks, we therefore propose a CMA-ES-based global forward-only learning framework. We further introduce structure-aware parameter decomposition and channel-specific credit assignment, factorizing the search into three subspaces. On BSDS500, our method reduces test MSE from 4.1783 to 4.0130 and improves PSNR from 42.621 to 42.787~dB over Adam/BP without backward propagation. Beyond these gains, training requires only 6 training-set coverages, compared with 1500 epochs for BP. Ablations show that structural decomposition and aligned credit assignment are more important than increasing population size, demonstrating the feasibility of whole-network forward-only learning for compact reconstruction networks. These results suggest a potential route toward forward-only optimization of compact reconstruction models for resource-constrained imaging and edge or near-sensor systems.
Oct 20, 2026

How Does Orthogonalization Adapt to the Neural Network Hessian Structure? A Gradient Self Outer-Product Analysis at Initialization

Muon orthogonalizes a weight matrix's momentum before each step, and on neural networks this simple preconditioner beats AdamW in a wide range of tasks. Existing analyses explain Muon within general optimization frameworks or on a few specific models, and many of them view orthogonalization as a preconditioner. Since the Hessian is the optimal preconditioner under a quadratic approximation of the loss, these works leave a basic question open: what makes orthogonalization suited to the Hessian of a neural network? Prior work shows that the layer-wise Hessian of a neural network is row-block diagonally dominant, and Muon's implicit preconditioner, the Kronecker product (VV⊤)1/2⊗I(\mathbf{V}\mathbf{V}^\top)^{1/2}\otimes\mathbf{I} with V\mathbf{V} the momentum, has the same row-block structure exactly when VV⊤\mathbf{V}\mathbf{V}^\top is diagonally dominant. Diagonal dominance of VV⊤\mathbf{V}\mathbf{V}^\top has been observed in language models, but why a neural network produces it has not been explained. This work provides a quantitative asymptotic theory for neural networks at initialization, and points out the intuition that the overparameterization of the weight matrices drives the diagonal dominance. We analyze E[GG⊤]\mathbb{E}[\mathbf{G}\mathbf{G}^\top], which equals E[VV⊤]\mathbb{E}[\mathbf{V}\mathbf{V}^\top] at initialization, in four canonical models: symmetric matrix factorization, deep linear networks, two-layer ReLU networks, and single-layer linear attention. In each case the diagonal entries outgrow the off-diagonal ones at explicit rates as the hidden dimensions grow; for the query, key and value weights of attention the off-diagonal vanishes altogether. Simulations of the four models and experiments on language models at initialization confirm the predicted scaling.
Oct 20, 2026

A General Topological Symmetry Index for Critical Orbits In Neural Network Loss Landscapes

A systematic understanding of the global structure of neural-network loss landscapes remains challenging due to their high dimensionality. Symmetries arising from network architectures and learning objectives provide a way to investigate aspects of this structure. Although a growing number of studies have drawn attention to these symmetries, a systematic characterization of them, as well as of the critical orbits they give rise to, remains incomplete. To address this gap, in this work we establish a topological symmetry index, based on the methodologies from the family of equivariant degree theories, to decode these symmetries. The frameworks apply to arbitrary network architectures and include methods applicable to degenerate critical orbits. The index detects the isotropy of the critical points and quantifies the corresponding net topological contributions. Several examples, including softplus MLPs, Transformers, and equivariant neural networks, are presented to illustrate the computations.
Oct 20, 2026

Do Neural Models Prefer Their Submodels?

A neural model is the set of functions that can be expressed by a given neural network architecture. Neural models contain natural submodels, such as those obtained by setting some neurons to be inactive. We study the geometry of neural models and ask whether they prefer their submodels. By this we mean whether local minimizers of the loss function lie in a given submodel. We introduce a geometric framework to determine if this occurs. For shallow multilayer perceptrons, we show that subnetworks with inactive hidden neurons are never preferred. For deep multilayer perceptrons, we give conditions under which subnetworks with inactive hidden neurons are not preferred. For lightning self-attention, lower-rank submodels are usually not preferred, but we identify examples where they are because they lie on a boundary of the model. Our results show that, aside from a few exceptions, preference for submodels cannot be explained by the geometry of the neural model, and must instead come from other sources such as training dynamics.
Oct 19, 2026

Dynamical Batching: How Class-Cycling Training Shapes the Internal Structure of Neural Networks

We investigate whether cyclically varying the class composition of training batches can modify both the performance and the internal organization of neural networks. Motivated by nonequilibrium systems, in which cycling between related sub-goals can make switching between their solutions increasingly efficient, we introduce Dynamical Batching (DB), a sampling protocol that periodically emphasizes each class while leaving the loss and the optimizer unchanged, and compare it with Uniform Batching (UB), in which every mini-batch has equal class proportions. In the regime where standard training fails, DB changes the outcome: on a three-arm spiral, a capacity-limited MLP that plateaus far below perfect accuracy under UB fits the training data under DB. Tracking individual weights, we find that under DB the change between class-specialized states is carried by a smaller set of weights than under UB, and that this set is markedly more persistent from cycle to cycle, on every synthetic dataset we test and in a wider network that fits the data under both protocols. This organization is a signature of the protocol rather than a predictor of its benefit: it appears equally where DB leaves accuracy unchanged. On CIFAR-10, the same protocol modestly improves the test accuracy of a channel-narrowed ResNet-10, a standard residual architecture, and of a second capacity-limited convolutional network. These results show that how class emphasis is distributed in time can shape both the solution a network reaches and the internal structure of the network.
Oct 17, 2026

Representations Have Trajectories: Characterizing Temporal Representation Dynamics in Neural Networks

Recent work has shown that deep neural networks exhibit an implicit low-rank simplicity bias. In this paper, we investigate the questions that naturally arise from this observation. What are the temporal consequences of low-rank simplicity bias during optimization? How does representation geometry evolve throughout optimization rather than only at convergence? Do different neural architectures exhibit distinct geometric optimization behaviors? To answer these questions, we conduct a large-scale empirical study tracking representational geometry throughout training. We find that neural networks follow three recurring geometric trajectories: persistent high-rank, collapse-and-recovery, and persistent low-rank. We further identify Destructive Representation Collapse (DRC) as a pathological regime in which representations collapse to rank-one and fail to recover, leading to a sustained loss of representational capacity. While collapse-and-recovery is consistently associated with successful optimization in controlled settings, experiments on real-world models show that the relationship between geometric trajectories and optimization outcomes is configuration dependent. Overall, our findings suggest that representation geometry evolves dynamically during training and provide a new perspective on how deep neural networks learn.