Introduction to the Attention Mechanism in Transformers

What Is Attention?

Attention is like when you read a sentence and your brain decides which words matter most for understanding its meaning.

Introduction to the attention mechanism in Transformers

Based on.

Scaled Dot-Product Attention

It is a mathematical formula that works like this:

1. Compare: It takes one word (the query) and compares it with all the other words (the keys)
2. Compute weights: It gives more importance to the most closely related words
3. Combine: It blends all the information according to those weights

Formula: Attention(Q,K,V) = softmax(QK^T/√dk)V

Multi-Head Attention

Instead of using a single attention “head,” it uses several (8 in this case):

• Each head focuses on different aspects of the relationships between words
• It is like having 8 different perspectives on the same text
• At the end, all the perspectives are combined

Three Uses in the Transformer:

1. Encoder-Decoder: The decoder looks at the entire input
2. Encoder self-attention: Each word looks at all the others in its layer
3. Decoder self-attention: Each word looks only at the previous ones (it cannot “cheat” by seeing the future)

In short: It is a mechanism that helps the model understand which parts of the text are most relevant to each word.

In the formula Attention(Q,K,V) = softmax(QK^T/√dk)V, the T stands for transpose.

What Is the Transpose?

The transpose of a matrix means flipping the matrix by swapping rows and columns.

Example:

Original matrix K:    Transpose K^T:
[1  2  3]            [1  4]
[4  5  6]            [2  5]
                     [3  6]

Why Is K^T Used?

To be able to perform matrix multiplication between Q and K:

• Q has dimensions: (number of queries × dk)
• K has dimensions: (number of keys × dk)
• K^T has dimensions: (dk × number of keys)

Therefore:

• Q × K^T = (number of queries × dk) × (dk × number of keys)
• Result = (number of queries × number of keys)

This gives us a matrix in which each element represents how compatible each query is with each key.

In Python code:

scores = torch.matmul(query, key.transpose(-2, -1))

The .transpose(-2, -1) is equivalent to applying the T (transpose).

dk is the dimension of the keys and queries.

Simple Definition

dk = the number of elements (features) in each key and query vector.

In the Context of the Transformer

According to the paper:

• dk = dv = dmodel/h = 64

Where:

• dmodel = total model dimension (for example, 512)
• h = number of attention heads (8 in this case)
• dk = 512 ÷ 8 = 64

Why Is It Important?

It is used in two key places:

1. In the division by √dk:

   scores = QK^T / √dk

• If dk = 64, then √dk = 8
• This normalizes the scores to keep them from becoming too large

2. In the matrix dimensions:
   • Each query vector (Q) has 64 elements
   • Each key vector (K) has 64 elements
   • Each value vector (V) also has 64 elements

Why Divide by √dk?

As the paper explains: when dk is large, the QK^T dot products become very large, causing the softmax function to have extremely small gradients. Dividing by √dk stabilizes training.

Example: If dk = 64, we divide by √64 = 8 to keep the values within a reasonable range.