Attention is Not All You Need: Pure Attention Loses Rank Doubly Exponentially with Depth

Abstract

Attention-based architectures have become ubiquitous in machine learning, yet our understanding of the reasons for their effectiveness remains limited. This work proposes a new way to understand self-attention networks: we show that their output can be decomposed into a sum of smaller terms, each involving the operation of a sequence of attention heads across layers. Using this decomposition, we prove that self-attention possesses a strong inductive bias towards "token uniformity". Specifically, without skip connections or multi-layer perceptrons (MLPs), the output converges doubly exponentially to a rank-1 matrix. On the other hand, skip connections and MLPs stop the output from degeneration. Our experiments verify the identified convergence phenomena on different variants of standard transformer architectures.

References in corpus (13)

• Language Models are Few-Shot Learners
• ALBERT: A Lite BERT for Self-supervised Learning of Language Representations
• Linformer: Self-Attention with Linear Complexity
• Universal Transformers
• Big Bird: Transformers for Longer Sequences
• Synthesizer: Rethinking Self-Attention in Transformer Models
• Lipschitz regularity of deep neural networks: analysis and efficient estimation
• Multi-Head Attention: Collaborate Instead of Concatenate
• On the Turing Completeness of Modern Neural Network Architectures
• Masked Language Modeling for Proteins via Linearly Scalable Long-Context Transformers
• The Lipschitz Constant of Self-Attention
• The Depth-to-Width Interplay in Self-Attention
• Lipschitz Networks and Distributional Robustness

Cited by in corpus (18)

• ViTGAN: Training GANs with Vision Transformers
• A Survey of Visual Transformers
• Refiner: Refining Self-attention for Vision Transformers
• SpectralFormer: Rethinking Hyperspectral Image Classification with Transformers
• MetaFormer Is Actually What You Need for Vision
• Analogous to Evolutionary Algorithm: Designing a Unified Sequence Model
• Not All Attention Is All You Need
• Can Vision Transformers Perform Convolution?
• On the Expressive Power of Self-Attention Matrices
• An Empirical Study: Extensive Deep Temporal Point Process
• Locally Enhanced Self-Attention: Combining Self-Attention and Convolution as Local and Context Terms
• GroupBERT: Enhanced Transformer Architecture with Efficient Grouped Structures
• Augmented Shortcuts for Vision Transformers
• Graph Conditioned Sparse-Attention for Improved Source Code Understanding
• Skyformer: Remodel Self-Attention with Gaussian Kernel and Nyström Method
• EfficientBERT: Progressively Searching Multilayer Perceptron via Warm-up Knowledge Distillation
• Multi-view 3D Reconstruction with Transformer
• Abstraction, Reasoning and Deep Learning: A Study of the "Look and Say" Sequence