Encoder: Text tokens -> Fixed representation vector

Decoder: Fixed representation vector + N decoded text tokens -> N+1th text token

Encoder/Decoder architecture: You take some tokenized text, run an encoder on it to get a fixed representation vector, and then recursively apply the decoder to your fixed representation vector and the 0...N tokens you've already produced to produce the N+1th token.

Decoder-only architecture: You take some tokenized text, and recursively apply a decoder to the 0...N tokens you've already produced to produce the N+1th token (without ever using an encoded representation vector).

Basically, an encoder produces this intermediate output which a decoder knows how to combine with some existing output to create more output (imagine, e.g., encoding a sentence in French, and then feeding a decoder the vector representation of that sentence plus the three words you've translated so far, so that it can figure out the next word in the translation). A decoder can be made to require an intermediate context vector, or (this is how it's done in decoder-only architectures) it can be made to require only the text produced so far.

Encoder in the T5 sense doesn't produce a fixed vector, it produces one encoded vector for every step of input and all of that is given to the decoder.

The only difference between encoder/decoder and decoder-only is masking:

In an encoder, none of the tokens are masked at any step, and are all visible in both directions to the encoder. Each output of the encoder can attend to any input of the encoder.

In the decoder, the tokens are masked causally - each N+1 token can only attend to the previous N tokens.

You can think of encoder/decoder models as specifically addressing the translation problem. They are also known as sequence-to-sequence models.

Take the task of translation. A translator needs to keep in mind the original text and the translation so far in order to predict the next translated token. The original text is encoded, and the translation so far is passed into the decoder to generate the next translated token. The next token is appended to the translation and the process repeats autoregressively.

Decoder-only models use just the decoder architecture of encoder/decoders. They are prompted and generate completions autoregressively.

Encoder-only models use just the encoder architecture which you can think of similarly to embedding. A task here is, producing vectors where vector distance is related to the semantic similarity of the input documents. This can be useful for retrieval tasks among other things.

You can of course translate using just the decoder, by constructing a "please translate this from A to B, <original text>" prompt and generating tokens just using the decoder. I'll leave it to people with more expertise than I do describe the pros and cons of these.

The biggest difference is when you feed a sequence into a decoder only model, it will only attend to previous tokens when computing hidden states for the current token. So the hidden states for the nth token is only based on tokens <n. This is where you hear the talk about "causal masking", as the attention matrix is masked to achieve this restriction. Encoder architectures on the other hand allow for each position in the sequence to attend to every other position in the sequence.

Encoder architectures have been used for semantic analysis, and feature extraction of sequences, and encoder only for generation (i.e. next token prediction).

Don't be embarrassed. This article makes the mistake of _saying_ they're going catch the under-informed up to speed but then immediately dives all the way in to the deep end.

The key to understanding the difference is that transformers are attention models where tokens can "attend" to different tokens.

Encoder models allow all tokens to attend to every other token. This increases the number of connections and makes it easier for the model to reason, but requires all tokens at once to produce any output. These models generally can't generate text.

Decoder models only allow tokens to attend to previous tokens in the sequence. This decreases the amount of tokens, but allows the model to be run incrementally, one token at a time. This incremental processing is key to allowing the models to generate text.

arent a lot of transformers built in a way where attention is only applied to previous tokens in sequence, even though its fully possible to apply it both ways?

That's the autoregressive aspect. The decoder aspect is that the last layer converts representations into output sequences (and the generation happens autoregressively, one at a time). Similarly at the last layer an encoder outputs a representation/embedding (while being able to attend to the entire sequence).

If you look at the classical [transformer architecture picture](https://en.wikipedia.org/wiki/Transformer_(deep_learning_arc...) there is an "encoder" tower on the left and a "decoder" tower on the right.

- Bert is encoder only.

- GPT is decoder only.

- T5 uses both the encoder and the decoder.