This chapter covers results in the paper Self-Attention Attribution: Interpreting Information Interactions Inside Transformer available at this Arxiv link 2021 version.
---
primaryColor: steelblue
shuffleQuestions: false
shuffleAnswers: true
---
### For an input of size n token, the dimension of the attention matrix is?
- [ ] n x 1.
- [x] n x n.
> Correct.
- [ ] 1 x n.
### Attention scores are not enough because
1. [x] they are too dense.
> Correct.
1. [ ] they are too sparse.
1. [ ] they are randomly generated.
1. [ ] they do not depend on the input.
Attributions using IG and Attention
---
primaryColor: steelblue
shuffleQuestions: false
shuffleAnswers: true
---
### Attributions are computed using a discrete approximation of the IG-style integral in this paper. The suggested number of samples (n) for the discrete approximation in the paper is:
- [ ] n = 3.
- [x] n = 20.
> Correct.
- [ ] n = 1000 or more.
Experiments using Attributions
---
primaryColor: steelblue
shuffleQuestions: false
shuffleAnswers: true
---
### The Pearson correlation between heads for similar tasks in the paper is
- [ ] close to 0.
- [x] close to 1.
> Correct.
- [ ] close to -1.
Conclusions
---
primaryColor: steelblue
shuffleQuestions: false
shuffleAnswers: true
---
### Self-attention attribution, as presented in this paper,
- [ ] does not rely on attention.
- [x] makes the self-attention mechanism more explainable.
> Correct.
- [ ] is independent of the input to the transformer.
Code and Assignment
There are no programming assignments for this lecture.