Imagine a spotlight that shines on the most important words in a sentence to help the computer understand the main idea. This mechanism helps the machine focus on what matters while ignoring irrelevant details.

The Spotlight of Attention

Computers cannot read text, so they turn every word into a long list of numbers called an embedding. These numbers represent the word's meaning, like coordinates on a giant map of ideas.

From Words to Numbers

Transformers read whole sentences at once, so they need 'positional encoding' to remember the sequence of words. This tells the computer that 'dog bites man' is different from 'man bites dog.'

Why Word Order Matters

The 'self-attention' trick allows a word to look at its neighbors to pick the right meaning. For example, it helps the computer know if 'bank' refers to a river or a place that holds money.

Understanding Word Context

Think of a 'query' as a search term and a 'key' as the label on a file folder. The computer compares them to find which words in a sentence have the most relevant information.

Queries and Keys

Once the computer finds a match between a query and a key, it extracts the 'value' or the actual content. This is the specific piece of knowledge the model uses to build its final answer.

The Value Information

Instead of one spotlight, 'multi-head attention' uses several at once to see different patterns. One head might focus on grammar, while another focuses on the emotions behind the words.

Looking at Many Angles

Unlike older AI that read one word at a time, Transformers process entire paragraphs simultaneously. This 'parallel processing' makes them much faster and better at handling long documents.

Processing Everything at Once

Transformers are great at remembering something mentioned at the very beginning of a long book. This ability to link far-apart ideas is called capturing 'long-range dependencies.'

Connecting Distant Ideas

The model uses a math tool called 'Softmax' to turn its search scores into percentages. It ensures that the most important words get the highest percentage of the computer's focus.

Turning Scores into Importance

Between every attention step, a 'feed-forward' network cleans up the information. It acts like an editor, polishing the raw data to make the word meanings even more accurate.

Refining the Raw Data

To prevent the AI from getting confused in deep layers, it uses 'residual connections' or shortcuts. These allow original information to skip ahead, keeping the learning process stable.

Information Shortcuts

At their core, these models are just advanced guessing machines that predict the most likely next word. By doing this billions of times, they eventually learn to speak like humans.

How Computers Pay Attention