As someone who is new to deep learning, I am only familiar with self-attention.
I'm designing a model. Imagine there are n data, which the $i_{th}$ data can be represented as a vector $x_i$. And the data has an attribute $a$, the attribute of the $i_{th}$ data is $a_i$. now I only use one query vector $q$ to multiply each data to calculate the result as the weight $w_i$ ($w_i=q^t * x_i$) , and use the weight of each data to multiply each data's attribute $a_i$ as the final result. ($A=\sum w_i * a_i$) Does this kind of attention exist? if so what its name is?
For example, $x_i$ is a 768-dimensional vector representing each sentence, and $s_i$ is a 3-dimensional vector representing each sentence's sentiment. $q$ is a 768-dimensional vector multiplied with each $x_i$ to produce each sentence's weight $w_i$. And the weighted sum $S=\sum w_i∗s_i$ is the overall sentiment of all sentences.
