39 lines
1.4 KiB
Markdown
39 lines
1.4 KiB
Markdown
|
- [[Attention|Self-attention]]
|
||
|
|
- Weighting significance of parts of the input
|
||
|
|
- Including recursive output
|
||
|
|
- Similar to [[RNN]]s
|
||
|
|
- Process sequential data
|
||
|
|
- Translation & text summarisation
|
||
|
|
- Differences
|
||
|
|
- Process input all at once
|
||
|
|
- Largely replaced [[LSTM]] and gated recurrent units (GRU) which had attention mechanics
|
||
|
|
- No recurrent structure
|
||
|
|
|
||
|
|
![[transformer-arch.png]]
|
||
|
|
|
||
|
|
## Examples
|
||
|
|
- BERT
|
||
|
|
- Bidirectional Encoder Representations from Transformers
|
||
|
|
- Google
|
||
|
|
- Original GPT
|
||
|
|
|
||
|
|
[transformers-explained-visually-part-1-overview-of-functionality](https://towardsdatascience.com/transformers-explained-visually-part-1-overview-of-functionality-95a6dd460452)
|
||
|
|
# Architecture
|
||
|
|
## Input
|
||
|
|
- Byte-pair encoding tokeniser
|
||
|
|
- Mapped via word embedding into vector
|
||
|
|
- Positional information added
|
||
|
|
|
||
|
|
## Encoder/Decoder
|
||
|
|
- Similar to seq2seq models
|
||
|
|
- Create internal representation
|
||
|
|
- Encoder layers
|
||
|
|
- Create encodings that contain information about which parts of input are relevant to each other
|
||
|
|
- Subsequent encoder layers receive previous encoding layers output
|
||
|
|
- Decoder layers
|
||
|
|
- Takes encodings and does opposite
|
||
|
|
- Uses incorporated textual information to produce output
|
||
|
|
- Has attention to draw information from output of previous decoders before drawing from encoders
|
||
|
|
- Both use [[attention]]
|
||
|
|
- Both use dense layers for additional processing of outputs
|
||
|
|
- Contain residual connections & layer norm steps
|