Dissecting the Interplay of Attention Paths in a Statistical Mechanics Theory of Transformers

• URL: http://arxiv.org/abs/2405.15926v1
• Date: Fri, 24 May 2024 20:34:18 GMT
• Title: Dissecting the Interplay of Attention Paths in a Statistical Mechanics Theory of Transformers
• Authors: Lorenzo Tiberi, Francesca Mignacco, Kazuki Irie, Haim Sompolinsky,
• Abstract summary: We consider a deep multi-head self-attention network, that is closely related to Transformers yet analytically tractable. We develop a statistical mechanics theory of Bayesian learning in this model. Experiments confirm our findings on both synthetic and real-world sequence classification tasks.
• Score: 14.59741397670484
• License: http://arxiv.org/licenses/nonexclusive-distrib/1.0/
• Abstract: Despite the remarkable empirical performance of Transformers, their theoretical understanding remains elusive. Here, we consider a deep multi-head self-attention network, that is closely related to Transformers yet analytically tractable. We develop a statistical mechanics theory of Bayesian learning in this model, deriving exact equations for the network's predictor statistics under the finite-width thermodynamic limit, i.e., $N,P\rightarrow\infty$, $P/N=\mathcal{O}(1)$, where $N$ is the network width and $P$ is the number of training examples. Our theory shows that the predictor statistics are expressed as a sum of independent kernels, each one pairing different 'attention paths', defined as information pathways through different attention heads across layers. The kernels are weighted according to a 'task-relevant kernel combination' mechanism that aligns the total kernel with the task labels. As a consequence, this interplay between attention paths enhances generalization performance. Experiments confirm our findings on both synthetic and real-world sequence classification tasks. Finally, our theory explicitly relates the kernel combination mechanism to properties of the learned weights, allowing for a qualitative transfer of its insights to models trained via gradient descent. As an illustration, we demonstrate an efficient size reduction of the network, by pruning those attention heads that are deemed less relevant by our theory.

Related papers

• A Theory for Compressibility of Graph Transformers for Transductive Learning [6.298115235439078]
  Transductive tasks on graphs differ fundamentally from typical supervised machine learning tasks. All train/test/validation samples are present during training, making them more akin to a semi-supervised task. We establish some theoretical bounds on how and under what conditions the hidden dimension of these networks can be compressed.
  arXiv  Detail & Related papers  (2024-11-20T04:20:17Z)
• Training Nonlinear Transformers for Chain-of-Thought Inference: A Theoretical Generalization Analysis [82.51626700527837]
  Chain-of-shift (CoT) is an efficient method that enables the reasoning ability of large language models by augmenting the query using examples with multiple intermediate steps. We show that despite the theoretical success of CoT, it fails to provide an accurate generalization when CoT does.
  arXiv  Detail & Related papers  (2024-10-03T03:12:51Z)
• 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)
• What Improves the Generalization of Graph Transformers? A Theoretical Dive into the Self-attention and Positional Encoding [67.59552859593985]
  Graph Transformers, which incorporate self-attention and positional encoding, have emerged as a powerful architecture for various graph learning tasks. This paper introduces first theoretical investigation of a shallow Graph Transformer for semi-supervised classification.
  arXiv  Detail & Related papers  (2024-06-04T05:30:16Z)
• Neural Tangent Kernels Motivate Graph Neural Networks with Cross-Covariance Graphs [94.44374472696272]
  We investigate NTKs and alignment in the context of graph neural networks (GNNs) Our results establish the theoretical guarantees on the optimality of the alignment for a two-layer GNN. These guarantees are characterized by the graph shift operator being a function of the cross-covariance between the input and the output data.
  arXiv  Detail & Related papers  (2023-10-16T19:54:21Z)
• Beyond spectral gap: The role of the topology in decentralized learning [58.48291921602417]
  In data-parallel optimization of machine learning models, workers collaborate to improve their estimates of the model. This paper aims to paint an accurate picture of sparsely-connected distributed optimization when workers share the same data distribution. Our theory matches empirical observations in deep learning, and accurately describes the relative merits of different graph topologies.
  arXiv  Detail & Related papers  (2022-06-07T08:19:06Z)
• How does unlabeled data improve generalization in self-training? A one-hidden-layer theoretical analysis [93.37576644429578]
  This work establishes the first theoretical analysis for the known iterative self-training paradigm. We prove the benefits of unlabeled data in both training convergence and generalization ability. Experiments from shallow neural networks to deep neural networks are also provided to justify the correctness of our established theoretical insights on self-training.
  arXiv  Detail & Related papers  (2022-01-21T02:16:52Z)
• The edge of chaos: quantum field theory and deep neural networks [0.0]
  We explicitly construct the quantum field theory corresponding to a general class of deep neural networks. We compute the loop corrections to the correlation function in a perturbative expansion in the ratio of depth $T$ to width $N$. Our analysis provides a first-principles approach to the rapidly emerging NN-QFT correspondence, and opens several interesting avenues to the study of criticality in deep neural networks.
  arXiv  Detail & Related papers  (2021-09-27T18:00:00Z)
• Spectral Bias and Task-Model Alignment Explain Generalization in Kernel Regression and Infinitely Wide Neural Networks [17.188280334580195]
  Generalization beyond a training dataset is a main goal of machine learning. Recent observations in deep neural networks contradict conventional wisdom from classical statistics. We show that more data may impair generalization when noisy or not expressible by the kernel.
  arXiv  Detail & Related papers  (2020-06-23T17:53:11Z)

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.