9 - Transformers

ucla | CS 162 | 2024-02-09 02:15

---

Table of Contents

• Seq2Seq Models
  • Encoder-Decoder Structure
  • Seq2Seq w/ Attention (Weighted sum)
• Transformers
  • Self-Attention
  • Multi-Headed Attention
  • Positional Embedding
  • Residual Connections and Layer Norm
  • Encoder-Decoder Attention

Seq2Seq Models

• input one sequence output another sequence
• e.g., audio -> text, text -> text (contextualized), text -> text

  Encoder-Decoder Structure

• idea is to use output of last cell as encoded vector bc (w/ RNN at least) the last cell contains information and contxt on the entire sentence so far

  Seq2Seq w/ Attention (Weighted sum)

• the pathway of encoder to context to attention vector; we compute attention by looking at all the attention vectors made by the encoder ($\overline{h_s}$) and compare to the decoder cell attention ($h_t$) (see arrows from sep token in decode to attention weights from the encoder)
• we make predictions based on attention vectors instead of hidden states

Transformers

• relies entirely on attention for encoder-decoder model instead of relying on recurrent dependency -> parallelizable

  Self-Attention

• attention between every token to every other token
• using QKV, compute the following circuits:
• QK circuit:
  • Generate $\hat{O}=Q{K}^T$ then norm & scale and softmax to get attentions $Z=\text{attention}=O=\text{softmax}(\frac{Q{K}^T}{\sqrt{d_k}})$
• OV circuit:
  • generate contextualized (encoded) vectors by multiplying attention to each key/value s.t. $C=OV=\text{softmax}(\frac{Q{K}^T}{\sqrt{d_k}})\cdot V$
• we can parallelize and introduce new weights by first making the QKV matrices using weights and the inputs: $Q=X\cdot W^Q$$K=X\cdot W^K$ $V=X\cdot W^V$

  Multi-Headed Attention

• we create multiple self-attention heads using multiple QKV weight matrices and generate multiple contextualized versions of the input vector from each head:
• then we concatenate the outputs and multiply by new output weight matrix $W^O$ and get a final contextualized matrix (by projection using the new weights) for the inputs:

  Positional Embedding

• something to be said abt positional relevance (words closer together likely are related in the same context), so we also include positional encoding along with input embeddings:
• t is the word position, k is the parity, i is the dimension index, d is the dimensionality

  Residual Connections and Layer Norm

• information loss of the input vector multiple steps in + vanishing gradients -> add back in the sublayer through a residual connection:
  • 
• and we want to normalize outputs by adding back in the original embeddings to the contextualized vectors
• for each dimension, normalize value wrt to mean to prevent vectors from becoming to large

  Encoder-Decoder Attention

• similar self attention but apply a look ahead mask after dimensional scaling in the self attention
• then pass that into Encoder-Decoder attention which is basically multi-headed attention but pass in Q,K as the contextualized representations from the Encoder