Drops recurrence entirely and stacks self-attention with feed-forward layers. Because no step waits on the previous one, the whole sequence trains in parallel, which is what let model size scale.
supersedescorrects · extends
classifiesspecializes · part-of
substitutes forapproximates · alternative-to
depends onrequires · validates
Colour is the family; a dashed line is the second member of it.
Drag to pan · scroll to zoom · click a node to open it
This node
correctsfixes a defect in LSTMrecurrence forbids parallelism across sequence positions
requiresdoes not work without Attentionit is the only operation left that mixes positions
requiresdoes not work without Layer Normalizationthe residual stack does not converge without it
Referenced by
specializesAutoregressive Language Model is a specific case of thisdecoder only, predicting forward with a causal mask
specializesBERT is a specific case of thisencoder only, trained by masking rather than predicting
correctsKV Cache fixes a defect in thisgeneration would recompute the prefix per token
extendsLoRA adds capability to thisadapts a frozen model through small low-rank updates
correctsMixture of Experts fixes a defect in thiscompute grows with capacity if every weight is used