When using Attention, for example with LSTM (but not necessarily), one can use the following methods to attend:

  1. MLP: $ug(W^1v+W^2q)$
  2. dot product: $v \cdot q$
  3. biaffine transform: $v^TWq$

($v$ is the attended vector which is used for prediction, $q$ is the query vector determining the weighted sum, used in the encoded input vectors attention)

What are the pros/cons of using each attention method? Or, what are the practical differences in the result?

  • $\begingroup$ You introduced the new tag [attention], can you please write a tag wiki? $\endgroup$ Feb 20, 2018 at 19:22

1 Answer 1


I think the main difference is just the amount of model capacity / complexity you want -- you'd probably start with dot product, and then pick increasingly elaborate methods if that doesn't fit properly.

Another consideration is the "type" of the vectors. If both $v$ and $q$ are word/sentence embeddings, a dot product seems straightforward, but what if $v$ is a sentence embedding and $q$ is the encoded form of an image? Then, taking the dot product makes less sense, since you are saying that the components of $v$ should somehow correspond to the same components in $q$.

Of course, this can also be a good thing if you are trying to come up with a single embedding space for multimodal inputs. So depending on this, you may or may not try to use biaffine attention, which doesn't assume $v$ and $q$ are the same type.

As for bi-affine vs MLP, note that bi-affine allows the easy modeling of quadratic interactions between $v$ and $q$, whereas MLP is more "linear". (See this related question on quadratic neurons in NTNs: What is the "expressive power" of the composition function in a Recursive Neural Tensor Network?)


Your Answer

By clicking “Post Your Answer”, you agree to our terms of service and acknowledge that you have read and understand our privacy policy and code of conduct.

Not the answer you're looking for? Browse other questions tagged or ask your own question.