Mind the Gap: a Spectral Analysis of Rank Collapse and Signal Propagation in Transformers

• URL: http://arxiv.org/abs/2410.07799v1
• Date: Thu, 10 Oct 2024 10:34:18 GMT
• Title: Mind the Gap: a Spectral Analysis of Rank Collapse and Signal Propagation in Transformers
• Authors: Alireza Naderi, Thiziri Nait Saada, Jared Tanner,
• Abstract summary: This paper examines signal propagation in textitattention-only transformers from a random matrix perspective. We show that a textitspectral gap between the two largest singular values of the attention matrix causes rank collapse in width. We propose a novel, yet simple, practical solution to resolve rank collapse in width by removing the spectral gap.
• Score: 3.686808512438363
• License: http://creativecommons.org/licenses/by/4.0/
• Abstract: Attention layers are the core component of transformers, the current state-of-the-art neural network architecture. However, \softmaxx-based attention puts transformers' trainability at risk. Even \textit{at initialisation}, the propagation of signals and gradients through the random network can be pathological, resulting in known issues such as (i) vanishing/exploding gradients and (ii) \textit{rank collapse}, i.e. when all tokens converge to a single representation \textit{with depth}. This paper examines signal propagation in \textit{attention-only} transformers from a random matrix perspective, illuminating the origin of such issues, as well as unveiling a new phenomenon -- (iii) rank collapse \textit{in width}. Modelling \softmaxx-based attention at initialisation with Random Markov matrices, our theoretical analysis reveals that a \textit{spectral gap} between the two largest singular values of the attention matrix causes (iii), which, in turn, exacerbates (i) and (ii). Building on this insight, we propose a novel, yet simple, practical solution to resolve rank collapse in width by removing the spectral gap. Moreover, we validate our findings and discuss the training benefits of the proposed fix through experiments that also motivate a revision of some of the default parameter scaling. Our attention model accurately describes the standard key-query attention in a single-layer transformer, making this work a significant first step towards a better understanding of the initialisation dynamics in the multi-layer case.

Related papers

• Unveiling Induction Heads: Provable Training Dynamics and Feature Learning in Transformers [54.20763128054692]
  We study how a two-attention-layer transformer is trained to perform ICL on $n$-gram Markov chain data. We prove that the gradient flow with respect to a cross-entropy ICL loss converges to a limiting model.
  arXiv  Detail & Related papers  (2024-09-09T18:10:26Z)
• Unveiling and Controlling Anomalous Attention Distribution in Transformers [8.456319173083315]
  Waiver phenomenon allows elements to absorb excess attention without affecting their contribution to information. In specific models, due to differences in positional encoding and attention patterns, we have found that the selection of waiver elements by the model can be categorized into two methods.
  arXiv  Detail & Related papers  (2024-06-26T11:53:35Z)
• Local to Global: Learning Dynamics and Effect of Initialization for Transformers [20.02103237675619]
  We focus on first-order Markov chains and single-layer transformers. We prove that transformer parameters trained on next-token prediction loss can either converge to global or local minima.
  arXiv  Detail & Related papers  (2024-06-05T08:57:41Z)
• Texture, Shape and Order Matter: A New Transformer Design for Sequential DeepFake Detection [57.100891917805086]
  Sequential DeepFake detection is an emerging task that predicts the manipulation sequence in order. This paper describes a new Transformer design, called TSOM, by exploring three perspectives: Texture, Shape, and Order of Manipulations.
  arXiv  Detail & Related papers  (2024-04-22T04:47:52Z)
• On the Convergence of Encoder-only Shallow Transformers [62.639819460956176]
  We build the global convergence theory of encoder-only shallow Transformers under a realistic setting. Our results can pave the way for a better understanding of modern Transformers, particularly on training dynamics.
  arXiv  Detail & Related papers  (2023-11-02T20:03:05Z)
• Signal Propagation in Transformers: Theoretical Perspectives and the Role of Rank Collapse [11.486545294602697]
  We shed new light on the causes and effects of rank collapse in Transformers. We show that rank collapse of the tokens' representations hinders training by causing the gradients of the queries and keys to vanish.
  arXiv  Detail & Related papers  (2022-06-07T09:07:24Z)
• Unraveling Attention via Convex Duality: Analysis and Interpretations of Vision Transformers [52.468311268601056]
  This paper analyzes attention through the lens of convex duality. We derive equivalent finite-dimensional convex problems that are interpretable and solvable to global optimality. We show how self-attention networks implicitly cluster the tokens, based on their latent similarity.
  arXiv  Detail & Related papers  (2022-05-17T04:01:15Z)
• Revisiting Over-smoothing in BERT from the Perspective of Graph [111.24636158179908]
  Recently over-smoothing phenomenon of Transformer-based models is observed in both vision and language fields. We find that layer normalization plays a key role in the over-smoothing issue of Transformer-based models. We consider hierarchical fusion strategies, which combine the representations from different layers adaptively to make the output more diverse.
  arXiv  Detail & Related papers  (2022-02-17T12:20:52Z)
• A Fourier-based Framework for Domain Generalization [82.54650565298418]
  Domain generalization aims at tackling this problem by learning transferable knowledge from multiple source domains in order to generalize to unseen target domains. This paper introduces a novel Fourier-based perspective for domain generalization. Experiments on three benchmarks have demonstrated that the proposed method is able to achieve state-of-the-arts performance for domain generalization.
  arXiv  Detail & Related papers  (2021-05-24T06:50:30Z)
• 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)

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.