paper-with-me

Papers

Transformers learn through gradual rank increase

2023-06-12 · NeurIPS 2023 11 · Enric Boix-Adsera, Etai Littwin, Emmanuel Abbe, Samy Bengio, Joshua Susskind

We identify incremental learning dynamics in transformers, where the difference between trained and initial weights progressively increases in rank. We rigorously prove this occurs under the simplifying assumptions of diagonal weight matrices and small initialization. Our experiments support the theory and also show that phenomenon can occur in practice without the simplifying assumptions.

📄 PDF Abstract BibTeX arXiv:2306.07042

Code (0)

등록된 구현이 없습니다.

Tasks

Incremental Learning

Similar Papers 제목 키워드 기반

InRank: Incremental Low-Rank Learning

2023-06-20 · Jiawei Zhao, Yifei Zhang, Beidi Chen, Florian Schäfer 외

The theory of greedy low-rank learning (GLRL) aims to explain the impressive generalization capabilities of deep learning. It proves that stochastic gradient-based training implicitly regularizes neural networks towards …

Computational Efficiency

Dissecting Lottery Ticket Transformers: Structural and Behavioral Study of Sparse Neural Machine Translation

2020-09-17 · EMNLP (BlackboxNLP) 2020 11 · Rajiv Movva, Jason Y. Zhao

Recent work on the lottery ticket hypothesis has produced highly sparse Transformers for NMT while maintaining BLEU. However, it is unclear how such pruning techniques affect a model's learned representations. By probing…

Machine TranslationNMTTranslation

Coarse-to-Fine Contrastive Learning on Graphs

2022-12-13 · Peiyao Zhao, Yuangang Pan, Xin Li, Xu Chen 외

Inspired by the impressive success of contrastive learning (CL), a variety of graph augmentation strategies have been employed to learn node representations in a self-supervised manner. Existing methods construct the con…

Contrastive LearningLearning-To-Rank

Dynamic Token-Pass Transformers for Semantic Segmentation

2023-08-03 · Yuang Liu, Qiang Zhou, Jing Wang, Fan Wang 외

Vision transformers (ViT) usually extract features via forwarding all the tokens in the self-attention layers from top to toe. In this paper, we introduce dynamic token-pass vision transformers (DoViT) for semantic segme…

SegmentationSemantic Segmentation

On the Role of Attention Masks and LayerNorm in Transformers

2024-05-29 · Xinyi Wu, Amir Ajorlou, Yifei Wang, Stefanie Jegelka 외

Self-attention is the key mechanism of transformers, which are the essential building blocks of modern foundation models. Recent studies have shown that pure self-attention suffers from an increasing degree of rank colla…