Attention: Weighing Input by Relative Importance

### The mental model

Attention is a mechanism that allows a neural network to dynamically weigh the importance of different parts of an input sequence. Instead of compressing an entire sequence into a single fixed-size vector, which can lose information, attention lets the model look back at the entire sequence and decide which parts are most relevant for the current step. It answers the question: "Given this part of the sequence, which other parts should I pay the most attention to?"

### How it works

The mechanism operates on sequences of vectors, known as token embeddings. For each token in the sequence, attention calculates a set of 'soft' weights. These weights represent the importance of every other token in the sequence relative to the current one. A higher weight means a stronger connection or relevance. The model then creates a new, context-aware representation for each token by taking a weighted sum of all token embeddings in the sequence, guided by these attention scores. This process effectively encodes the context from the entire sequence into each token's representation.

### When to use it

- **Natural Language Processing:** It's a core component in modern NLP models for tasks where context between words is critical. - **Encoding Long Sequences:** It is used to encode context across sequences of token embeddings, which can range in size from tens to millions of tokens. - **Relational Tasks:** Any task where the relationship between different elements in a sequence determines the outcome.

### When NOT to use it

- **Independent Inputs:** For problems where each part of the input sequence has equal and independent importance, and context between elements is irrelevant.

### One canonical example

Consider a model processing the sentence, "The gray cat sat on the mat." The attention mechanism helps the model understand relationships. When focusing on the word "sat," the model might assign high attention weights to "cat" (who sat) and "mat" (where it sat). This is done by calculating these 'soft' weights for the token embeddings of each word in the sentence relative to "sat." The resulting representation for "sat" is now enriched with information about the subject and location, all learned dynamically from the sequence itself.