Feature learning is widely regarded as the key mechanism distinguishing neural networks from fixed-kernel methods, yet its impact on the induced function space remains poorly understood. In this work, we precisely characterize how the function space spanned by the features of a two-layer neural network evolves during gradient descent training. We prove that, in the high-dimensional proportional regime, after a large gradient step the post-update feature distribution is well approximated by a target-dependent spiked Gaussian covariance. This induces a data-adaptive kernel that reshapes the function space and modifies its spectral structure. Our analysis reveals that feature learning can be interpreted as a distributional transformation in either parameter space or input space, equivalently as the introduction of a target-dependent kernel. In particular, it selectively amplifies eigenvalues aligned with the target direction and mixes leading eigenfunctions, coupling the top radial mode with a target-aligned quadratic harmonic. Overall, our results provide a precise function-space perspective on early-stage feature learning: rather than just rescaling a fixed kernel, gradient descent induces a data-adaptive deformation that preferentially enhances directions aligned with the signal in the data.

Stochastic gradient descent (SGD) has been studied extensively over the past decades due to its simplicity and broad applicability in machine learning. In this work, we analyze the local behavior of gradient descent and stochastic gradient descent for minimizing $C^2$-functions that satisfy the Polyak-Lojasiewicz (PL) inequality and under a multiplicative gradient noise model motivated by overparameterized neural networks. Using a geometric interpretation of the PL-condition, we prove a simple yet surprising fact: in this possibly non-convex setting, the asymptotic convergence rate of (S)GD matches the rate obtained for strongly convex quadratics.

We establish the first population risk bounds for Kolmogorov-Arnold Networks (KANs) trained by mini-batch SGD with gradient clipping, covering non-private SGD as well as differentially private SGD (DP-SGD) with Gaussian perturbations that interpolate between independent and temporally correlated noise. This setting is substantially closer to practice than prior KAN theory along two axes: training is by mini-batch SGD, the standard recipe for modern networks, rather than full-batch gradient descent (GD); and correlated-noise mechanisms have empirically shown a more favorable privacy-utility tradeoff than independent-noise mechanisms. Our results cover the corresponding full-batch GD and independent-noise DP-GD results for KANs by Wang et al. (2026), while yielding sharper fixed-second-layer specializations. The technical core is a new analysis route for correlated-noise DP training in the non-convex regime. Temporal dependence breaks the conditional-centering structure underlying standard one-step SGD arguments, and the projection step obstructs the exact cancellation structure of correlated perturbations. We address these difficulties through an auxiliary unprojected dynamics, a shifted iterate that absorbs the current noise perturbation, and a high-probability bootstrap certifying projection inactivity. Combining this optimization analysis with a stability-based generalization argument yields the stated population risk bounds. To the best of our knowledge, this is the first optimization and population risk analysis of a correlated-noise mechanism for DP training beyond convex learning, in particular for neural networks.

Asynchronous stochastic gradient descent (ASGD) is a standard way to exploit heterogeneous compute resources in distributed learning: instead of forcing fast workers to wait for slow ones, the server updates the model whenever a gradient arrives. Vanilla ASGD applies each arriving gradient with the same weight. When local data distributions are heterogeneous, this becomes problematic: faster workers contribute more updates, and we show theoretically that the method is biased toward a frequency-weighted average of the local objectives rather than the desired global objective. Existing remedies typically move away from the simple ASGD template by introducing gathering phases, buffering, or extra memory. We show that this is unnecessary. Keeping the standard ASGD mechanism, we recover the correct objective by rescaling worker-specific stepsizes in proportion to their computation times, so that each worker contributes the same aggregate learning rate over a cycle. In the non-convex setting, under smoothness and bounded heterogeneity assumptions, we prove that the resulting method, Rescaled ASGD, converges to stationary points of the correct global objective in the fixed-computation model. Its time complexity matches the known lower bound in the leading term, while the effects of staleness and data heterogeneity appear only in lower-order terms. Experiments confirm that the method converges to the correct objective and is competitive with state-of-the-art baselines.

Stein Variational Gradient Descent (SVGD), introduced in [LW16], is a deterministic interactingparticle method for sampling from a target probability measure π e V, only requiring access to V. In the mean-field and continuous-time limit, the distribution of particles converges to a flow (ρt) in the space of probability measures that solves a variant of the Fokker-Planck equation with a velocity field smoothed by weighted convolution with a positive definite kernel [LLN19]. This flow can be interpreted as the gradient flow of the relative entropy H( |π) with respect to a "kernelized" Wasserstein metric [Liu17]. The goal of this paper is to investigate the convergence of (ρt) towards π. To this end, we focus on the model case of Riesz kernels of order s on the d-dimensional torus Td. This is a family of translation-invariant kernels whose Fourier coefficients decay as |ξ| 2s. The parameter s hence directly controls the "smoothing strength" of the interaction; in particular, continuous kernels correspond to s > d/2, C1 kernels to s > (d+1)/2, and C2 kernels to s > (d+2)/2. What is known: qualitative weak convergence The starting point of convergence analyses is the energy dissipation formula [Liu17] d dt H(ρt|π) = Is(ρt|π), (1.1) Authors are listed in alphabetical order.

Neural networks are trained by minimizing loss functions with gradient-based optimizers. Cohen et al. [2021] observed that full-batch gradient descent operates at the edge of stability (EoS): the largest eigenvalue of the Hessian, called the sharpness, first rises (a phase called progressive sharpening) and then hovers at the stability threshold 2/η where η is the learning rate. Cohen et al. [2022] extended this picture to momentum methods and adaptive gradient methods, showing that each optimizer exhibits its own edge of stability. Rather than hovering at 2/η, the relevant quantity--the preconditioned sharpness--hovers at a hyperparameter-dependent threshold that depends on the optimizer (Table 2). In practice, the dominant optimizer in machine learning is Adam [Kingma and Ba, 2015], which differs from gradient descent in two respects.

Online learning from a stream of data is a defining feature of intelligence, yet modern machine learning systems often struggle in this setting, especially under distributional shift. To understand its basic properties, we study the relationship between online and offline learning in the context of kernel regression. We derive a closed-form expression for the function learned by online kernel regression, revealing that online kernel regression is equivalent to offline regression with shifted, inaccurate target outputs. Conversely, we show that by compensating for this effective shift in the teaching signal through target correction, online kernel-based learning can provably learn the same predictor as its offline counterpart. We derive both a closed-form expression for this target correction and an iterative form that can be applied sequentially. Applying this framework to image classification tasks on CIFAR-10 and CORe50, we show that online stochastic gradient descent with iteratively corrected targets outperforms learning with the true targets in continual learning settings. This work therefore provides a basic framework for analyzing and improving online learning in non-stationary environments.

We study a class of bilevel optimization problems in which both the upper- and lower-level problems have minimax structures. This setting captures a broad range of emerging applications. Despite the extensive literature on bilevel optimization and minimax optimization separately, existing methods mainly focus on bilevel optimization with lower-level minimization problems, often under strong convexity assumptions, and are not directly applicable to the minimax lower-level setting considered here. To address this gap, we develop penalty-based first-order methods for bilevel minimax optimization without requiring strong convexity of the lower-level problem. In the deterministic setting, we establish that the proposed method finds an $ε$-KKT point with $\tilde{O}(ε^{-4})$ oracle complexity. We further show that bilevel problems with convex constrained lower-level minimization can be reformulated as special cases of our framework via Lagrangian duality, leading to an $\tilde{O}(ε^{-4})$ complexity bound that improves upon the existing $\tilde{O}(ε^{-7})$ result. Finally, we extend our approach to the stochastic setting, where only stochastic gradient oracles are available, and prove that the proposed stochastic method finds a nearly $ε$-KKT point with $\tilde{O}(ε^{-9})$ oracle complexity.

One widely recognized interpretation for their empirical success is their ability to perform in-context learning (ICL): pretrained transformers are capable of performing previously unseen tasks based on demonstrations and examples in the prompt, without requiring any additional task-specific fine-tuning (Brown et al., 2020). A line of recent works interpret the in-context learning (ICL) capability of transformers from an algorithmic perspective, viewing transformers as models that can implicitly execute certain learning algorithms on the context examples. Specifically, Garg et al. (2022) proposes a theoretical framework for ICL in terms of learning a hypothesis class, and empirically shows that transformers can in-context learn the linear function class. Motivated by this empirical finding, several recent works attempt to theoretically study how transformers perform in-context learning on linear regression tasks. Aky urek et al. (2022); Von Oswald et al. (2023) construct multi-layer transformers with linear attention that can execute gradient descent on the an "in-context loss" defined on the context data, thereby enabling in-context learning of linear regression.

Decentralized learning is a learning process in which data is distributed across computational agents or collected by individual agents, and model parameters are computed as the consensus of the agents. It has gained a lot of interest for applications where agents can collaboratively learn a predictive model without sharing their own data, but sharing only their local models with their immediate neighbors to generate a global model [He et al., 2018, Hendrikx et al., 2019, Arjevani et al., 2020]. We assume there are N agents who are connected over an undirected communication network G = (V,E) where V = {1,...,N} represents the agents and E V V denotes the set of edges; i.e., if agent i and j are connected then (i,j) E implies (j,i) E. Suppose we have a collection of n independent and identically distributed (i.i.d.) data pairs zi = (ai,yi), where ai Rp is the feature vector and yi the label or response of the i-th observation. Let Z = [z1,z2,,zn] Rnp be sampled from the distribution p(Z|x) where the parameter x Rd has a common prior. The goal is to sample from the posterior distribution p(x|Z) p(Z|x)p(x) by distributing Z among N agents such that Zi = {zi1,zi2,,zini} is the subset of data exclusive to agent i.