Training Infinitely Deep and Wide Transformers

• URL: http://arxiv.org/abs/2605.17660v1
• Date: Sun, 17 May 2026 21:30:13 GMT
• Title: Training Infinitely Deep and Wide Transformers
• Abstract summary: This paper develops a rigorous framework for analyzing gradient-based training of transformers in the mean-field regime.<n>We prove the existence and uniqueness of gradient flow curves in the conditional Wasserstein metric space.<n>A key technical contribution is providing necessary and sufficient conditions for injectivity of the Neural Tangent Kernel.
• Score: 27.784304430081622
• License: http://creativecommons.org/licenses/by/4.0/
• Abstract: Transformers have become the dominant architecture in modern machine learning, yet the theoretical understanding of their training dynamics remains limited. This paper develops a rigorous mathematical framework for analyzing gradient-based training of transformers in the mean-field regime, where both the depth (number of layers) and width (number of attention heads) tend to infinity. While ResNet training can be understood as controlling a neural ODE, transformer training corresponds to controlling a neural PDE, due to the coupling of multiple token distributions through the attention mechanism. Our mean-field model features two types of measure representations: token distributions evolving through layers and attention parameters at each layer. We establish well-posedness of the forward pass through infinitely deep transformers, characterizing token evolution via flow maps that satisfy ODEs in function spaces. Using adjoint sensitivity analysis, we derive an explicit formula for the conditional Wasserstein gradient of the training risk, involving adjoint variables governed by backward ODEs. We prove the existence and uniqueness of gradient flow curves in the conditional Wasserstein metric space, establishing a rigorous foundation for gradient-based transformer training. A key technical contribution is providing necessary and sufficient conditions for injectivity of the Neural Tangent Kernel (NTK) for attention mechanisms: we show that NTK injectivity is equivalent to linear independence of log-sum-exp functions modulo affine functions, a condition satisfied by diverse token distributions, including discrete distributions, uniform distributions, and Gaussian mixtures. Under this NTK injectivity assumption, we prove that gradient flow converges to global minima when the initial loss is sufficiently small, eliminating spurious local minima from the optimization landscape.

Related papers

• Fluid-DiT: Graph-Free Diffusion Transformers for Fluid Flow Simulations Learning [50.053186229257555]
  We propose a Graph-Free Diffusion Transformer that replaces graph message passing with attention-based denoising.<n>By leveraging the transformer's global receptive field, Fluid-DiT naturally captures both local flow structures and long-range correlations.<n>It consistently outperforms graph-based diffusion baselines in both sample quality and distributional accuracy.
  arXiv  Detail & Related papers  (2026-08-07T12:28:45Z)
• Physics-informed neural particle flow for the Bayesian update step [0.8220217498103312]
  We propose a physics-informed neural particle flow, which is an amortized inference framework.<n>By embedding a governing partial differential equation (PDE) into the loss function, we train a neural network to approximate the transport velocity field.<n>We demonstrate that the neural parameterization acts as an implicit regularizer, mitigating the stiffness inherent to analytic flows.
  arXiv  Detail & Related papers  (2026-02-26T15:10:45Z)
• Flow matching Operators for Residual-Augmented Probabilistic Learning of Partial Differential Equations [0.5729426778193397]
  We formulate flow matching in an infinite-dimensional function space to learn a probabilistic transport.<n>We develop a conditional neural operator architecture based on feature-wise linear modulation for flow matching vector fields.<n>We show that the proposed method can accurately learn solution operators across different resolutions and fidelities.
  arXiv  Detail & Related papers  (2025-12-14T16:06:10Z)
• Layer-wise Quantization for Quantized Optimistic Dual Averaging [75.4148236967503]
  We develop a general layer-wise quantization framework with tight variance and code-length bounds, adapting to the heterogeneities over the course of training.<n>We propose a novel Quantized Optimistic Dual Averaging (QODA) algorithm with adaptive learning rates, which achieves competitive convergence rates for monotone VIs.
  arXiv  Detail & Related papers  (2025-05-20T13:53:58Z)
• Ultra-fast feature learning for the training of two-layer neural networks in the two-timescale regime [26.47265060394168]
  We study the convergence of methods for the training of mean-field-hidden-layer neural networks.<n>We show that a strategy enables provable convergence rates for the sampling of a teacher feature distribution.
  arXiv  Detail & Related papers  (2025-04-25T09:40:10Z)
• Training Dynamics of Transformers to Recognize Word Co-occurrence via Gradient Flow Analysis [97.54180451650122]
  We study the dynamics of training a shallow transformer on a task of recognizing co-occurrence of two designated words. We analyze the gradient flow dynamics of simultaneously training three attention matrices and a linear layer. We prove a novel property of the gradient flow, termed textitautomatic balancing of gradients, which enables the loss values of different samples to decrease almost at the same rate and further facilitates the proof of near minimum training loss.
  arXiv  Detail & Related papers  (2024-10-12T17:50:58Z)
• Non-asymptotic Convergence of Training Transformers for Next-token Prediction [48.9399496805422]
  Transformers have achieved extraordinary success in modern machine learning due to their excellent ability to handle sequential data. This paper provides a fine-grained non-asymptotic analysis of the training dynamics of a one-layer transformer. We show that the trained transformer presents non-token prediction ability with dataset shift.
  arXiv  Detail & Related papers  (2024-09-25T20:22:06Z)
• Adaptive Federated Learning Over the Air [108.62635460744109]
  We propose a federated version of adaptive gradient methods, particularly AdaGrad and Adam, within the framework of over-the-air model training. Our analysis shows that the AdaGrad-based training algorithm converges to a stationary point at the rate of $mathcalO( ln(T) / T 1 - frac1alpha ).
  arXiv  Detail & Related papers  (2024-03-11T09:10:37Z)
• Machine learning in and out of equilibrium [58.88325379746631]
  Our study uses a Fokker-Planck approach, adapted from statistical physics, to explore these parallels. We focus in particular on the stationary state of the system in the long-time limit, which in conventional SGD is out of equilibrium. We propose a new variation of Langevin dynamics (SGLD) that harnesses without replacement minibatching.
  arXiv  Detail & Related papers  (2023-06-06T09:12:49Z)
• Effects of Parameter Norm Growth During Transformer Training: Inductive Bias from Gradient Descent [44.44543743806831]
  We study the tendency for transformer parameters to grow in magnitude while saturated between these norms during training. As the parameters grow in magnitude, we prove that the network approximates a discretized network with saturated activation functions. Our results suggest saturation is a new characterization of an inductive bias implicit in GD of particular interest for NLP.
  arXiv  Detail & Related papers  (2020-10-19T17:40:38Z)
• Kernel and Rich Regimes in Overparametrized Models [69.40899443842443]
  We show that gradient descent on overparametrized multilayer networks can induce rich implicit biases that are not RKHS norms. We also demonstrate this transition empirically for more complex matrix factorization models and multilayer non-linear networks.
  arXiv  Detail & Related papers  (2020-02-20T15:43:02Z)
• Implicit Bias of Gradient Descent for Wide Two-layer Neural Networks Trained with the Logistic Loss [0.0]
  Neural networks trained to minimize the logistic (a.k.a. cross-entropy) loss with gradient-based methods are observed to perform well in many supervised classification tasks. We analyze the training and generalization behavior of infinitely wide two-layer neural networks with homogeneous activations.
  arXiv  Detail & Related papers  (2020-02-11T15:42:09Z)

This list is automatically generated from the titles and abstracts of the papers in this site.

This site does not guarantee the quality of this site (including all information) and is not responsible for any consequences.