Explore a comprehensive lecture on attention mechanisms in neural networks for natural language processing. Delve into the fundamentals of attention, including what to attend to, improvements to attention techniques, and specialized attention varieties. Examine a detailed case study on the "Attention is All You Need" paper. Learn about encoder-decoder models, sentence representations, and the basic idea behind attention as proposed by Bahdanau et al. in 2015. Discover various attention score functions, hierarchical structures, and techniques for handling multiple sources. Address common problems in neural models, such as dropping or repeating information, and explore solutions like incorporating Markov properties and bidirectional training. Gain insights into hard attention, the Transformer architecture, and various attention tricks. Conclude with training techniques, including masking for efficient training operations.
Overview
Syllabus
Intro
Encoder-decoder Models
Sentence Representations
Basic Idea (Bahdanau et al. 2015)
Calculating Attention (2)
A Graphical Example
Attention Score Functions (1)
Input Sentence
Hierarchical Structures (Yang et al. 2016)
Multiple Sources
Coverage • Problem: Neural models tends to drop or repeat
Incorporating Markov Properties (Cohn et al. 2015)
در Bidirectional Training
Hard Attention
Summary of the Transformer (Vaswani et al. 2017)
Attention Tricks
Training Tricks
Masking for Training . We want to perform training in as few operations as
Taught by
Graham Neubig
