Is the input/output of these models any different? Are they all just "text context goes in, scores for all tokens in the vocabulary come out" ? Is the difference only in how they achieve this output?
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.
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.
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.
Encoder architectures have been used for semantic analysis, and feature extraction of sequences, and encoder only for generation (i.e. next token prediction).
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.
The term for models that look only at previous tokens in the sequence is auto-regressive.
Encoder and decoder has nothing to do with this.
- Bert is encoder only.
- GPT is decoder only.
- T5 uses both the encoder and the decoder.