Generalized Linear Mode Connectivity for Transformers

• URL: http://arxiv.org/abs/2506.22712v1
• Date: Sat, 28 Jun 2025 01:46:36 GMT
• Title: Generalized Linear Mode Connectivity for Transformers
• Authors: Alexander Theus, Alessandro Cabodi, Sotiris Anagnostidis, Antonio Orvieto, Sidak Pal Singh, Valentina Boeva,
• Abstract summary: A striking phenomenon is linear mode connectivity (LMC), where independently trained models can be connected by low- or zero-loss paths.<n>Prior work has predominantly focused on neuron re-ordering through permutations, but such approaches are limited in scope.<n>We introduce a unified framework that captures four symmetry classes: permutations, semi-permutations, transformations, and general invertible maps.<n>This generalization enables, for the first time, the discovery of low- and zero-barrier linear paths between independently trained Vision Transformers and GPT-2 models.
• Score: 87.32299363530996
• License: http://arxiv.org/licenses/nonexclusive-distrib/1.0/
• Abstract: Understanding the geometry of neural network loss landscapes is a central question in deep learning, with implications for generalization and optimization. A striking phenomenon is linear mode connectivity (LMC), where independently trained models can be connected by low- or zero-loss paths, despite appearing to lie in separate loss basins. However, this is often obscured by symmetries in parameter space -- such as neuron permutations -- which make functionally equivalent models appear dissimilar. Prior work has predominantly focused on neuron re-ordering through permutations, but such approaches are limited in scope and fail to capture the richer symmetries exhibited by modern architectures such as Transformers. In this work, we introduce a unified framework that captures four symmetry classes: permutations, semi-permutations, orthogonal transformations, and general invertible maps -- broadening the set of valid reparameterizations and subsuming many previous approaches as special cases. Crucially, this generalization enables, for the first time, the discovery of low- and zero-barrier linear interpolation paths between independently trained Vision Transformers and GPT-2 models. These results reveal deeper structure in the loss landscape and underscore the importance of symmetry-aware analysis for understanding model space geometry.

Related papers

• Flow Equivariant Recurrent Neural Networks [2.900810893770134]
  In machine learning, neural network architectures that respect symmetries of their data are called equivariant.<n>We extend equivariant network theory to this regime of flows', capturing natural transformations over time.<n>We show that these models significantly outperform their non-equivariant counterparts in terms of training speed, length generalization, and velocity generalization.
  arXiv  Detail & Related papers  (2025-07-20T02:52:21Z)
• Symmetry in Neural Network Parameter Spaces [32.732734207891745]
  A significant portion of redundancy is explained by symmetries in the parameter space--transformations that leave the network function unchanged.<n>These symmetries shape the loss landscape and constrain learning dynamics, offering a new lens for understanding optimization, generalization, and model complexity.<n>We summarize existing literature, uncover connections between symmetry and learning theory, and identify gaps and opportunities in this emerging field.
  arXiv  Detail & Related papers  (2025-06-16T00:59:12Z)
• Relative Representations: Topological and Geometric Perspectives [53.88896255693922]
  Relative representations are an established approach to zero-shot model stitching.<n>We introduce a normalization procedure in the relative transformation, resulting in invariance to non-isotropic rescalings and permutations.<n>Second, we propose to deploy topological densification when fine-tuning relative representations, a topological regularization loss encouraging clustering within classes.
  arXiv  Detail & Related papers  (2024-09-17T08:09:22Z)
• EqNIO: Subequivariant Neural Inertial Odometry [33.96552018734359]
  We show that IMU data transforms equivariantly, when rotated around the gravity vector and reflected with respect to arbitrary planes parallel to gravity. We then map the IMU data into this frame, thereby achieving an invariant canonicalization that can be directly used with off-the-shelf inertial odometry networks.
  arXiv  Detail & Related papers  (2024-08-12T17:42:46Z)
• Curve Your Attention: Mixed-Curvature Transformers for Graph Representation Learning [77.1421343649344]
  We propose a generalization of Transformers towards operating entirely on the product of constant curvature spaces. We also provide a kernelized approach to non-Euclidean attention, which enables our model to run in time and memory cost linear to the number of nodes and edges.
  arXiv  Detail & Related papers  (2023-09-08T02:44:37Z)
• Regularization, early-stopping and dreaming: a Hopfield-like setup to address generalization and overfitting [0.0]
  We look for optimal network parameters by applying a gradient descent over a regularized loss function. Within this framework, the optimal neuron-interaction matrices correspond to Hebbian kernels revised by a reiterated unlearning protocol.
  arXiv  Detail & Related papers  (2023-08-01T15:04:30Z)
• Implicit Balancing and Regularization: Generalization and Convergence Guarantees for Overparameterized Asymmetric Matrix Sensing [28.77440901439686]
  A series of recent papers have begun to generalize this role for non-random Positive Semi-Defin (PSD) matrix sensing problems. In this paper, we show that the trajectory of the gradient descent from small random measurements moves towards solutions that are both globally well.
  arXiv  Detail & Related papers  (2023-03-24T19:05:52Z)
• Oracle-Preserving Latent Flows [58.720142291102135]
  We develop a methodology for the simultaneous discovery of multiple nontrivial continuous symmetries across an entire labelled dataset. The symmetry transformations and the corresponding generators are modeled with fully connected neural networks trained with a specially constructed loss function. The two new elements in this work are the use of a reduced-dimensionality latent space and the generalization to transformations invariant with respect to high-dimensional oracles.
  arXiv  Detail & Related papers  (2023-02-02T00:13:32Z)
• Optimizing Mode Connectivity via Neuron Alignment [84.26606622400423]
  Empirically, the local minima of loss functions can be connected by a learned curve in model space along which the loss remains nearly constant. We propose a more general framework to investigate effect of symmetry on landscape connectivity by accounting for the weight permutations of networks being connected.
  arXiv  Detail & Related papers  (2020-09-05T02:25:23Z)
• Masked Language Modeling for Proteins via Linearly Scalable Long-Context Transformers [42.93754828584075]
  We present a new Transformer architecture, Performer, based on Fast Attention Via Orthogonal Random features (FAVOR) Our mechanism scales linearly rather than quadratically in the number of tokens in the sequence, is characterized by sub-quadratic space complexity and does not incorporate any sparsity pattern priors. It provides strong theoretical guarantees: unbiased estimation of the attention matrix and uniform convergence.
  arXiv  Detail & Related papers  (2020-06-05T17:09:16Z)
• Inverse Learning of Symmetries [71.62109774068064]
  We learn the symmetry transformation with a model consisting of two latent subspaces. Our approach is based on the deep information bottleneck in combination with a continuous mutual information regulariser. Our model outperforms state-of-the-art methods on artificial and molecular datasets.
  arXiv  Detail & Related papers  (2020-02-07T13:48:52Z)

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.