I understand the transformer architecture (from "Attention is All You Need"), as well as how the attention is computed in the multi-headed attention layers.

What I'm confused on is why the output of an attention layer is a context vector. That is to say: what is it about the way that a transformer is trained causes the attention layers to learn context? What I would expect to see in the paper is a justification along the lines of "when you train a transformer using attention on sequence-to-sequence tasks, the attention layers learn context and here's why...". I believe it because I've seen the heatmaps that show that attention between related words, but I want to understand why that is necessarily the result of training a transformer.

Why couldn't it be the case that the attention layers learn some other features that happen to also be beneficial in sequence to sequence tasks? How do we know that they learn context, other than that's what we observe?

Again, I get the math and I know there are several posts about it. What I want to know is what about the math or the training process implies that the attention layers learn context.

  • 1
    $\begingroup$ what is your definition of the context here? In the most general sense context is just any kind of information that can help you to interpret something in a way it was to supposed to be interpreted. why that is necessarily the result of training a transformer - to that point, what else can it be result of? $\endgroup$
    – Deil
    Commented May 25, 2021 at 13:54
  • $\begingroup$ Awesome. Your second sentence: "In the most general sense context is just any kind of information that can help you to interpret something in a way it was to supposed to be interpreted" really gets at the core of the thing I was stuck on. $\endgroup$ Commented Jun 8, 2021 at 16:31

1 Answer 1


To provide a simplistic and less mathematical reasons. You can assume like this:

In a simple feed-forward neural network (a black-box of course), you shall learn the set of weights, learning a function to map inputs to outputs.

But, in the transformers based architecture, you have Attentions. Here, the weights are structured into Query, Key and Value (Q,K,V). These 3 set of weights drivers of attention and are responsible to learn the context. How precisely it works still remains a black-box like feed forward networks. But yeah, it works something like this, every token's embedding is transformed its query, key and value vectors using their respective weight matrices. For a given token, its query vector is multiplied with all the other tokens' key vector to obtain a value vector. These values determines the importance of every token with respect to the query token. Thus, with back-propagation, you try to optimize these Q, K, V weights, and thus learn it to better map the relationship between tokens.

  • $\begingroup$ Thanks for your response. But I'm still confused what it is about the sequence-to-sequence task causes the weights to learn a better relationship between the tokens and not something else. For example, in a classification task, the loss function is calculated based on the classification accuracy, so backpropagation naturally results in better classification accuracy. But the loss in sequence-to-sequence is not based strictly on attention layers learning context, so how do we know that's what they learn? $\endgroup$ Commented Nov 13, 2020 at 15:18
  • $\begingroup$ It's not just sequence-to-sequence task. Original transformers were trained on Seq-to-Seq task. But models like BERT (which is completely based on transformers), are learnt on masked language model objective. A different training objective. $\endgroup$ Commented Nov 13, 2020 at 17:34
  • $\begingroup$ Right, but the question still stands. What about these objectives cause the attention layers to learn context and not something else beneficial to the task? $\endgroup$ Commented Nov 13, 2020 at 17:46
  • $\begingroup$ They would still learn something beneficial to the task. It's not just context. $\endgroup$ Commented Nov 13, 2020 at 17:58

Your Answer

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

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