แนวคิดของแอตเทนชัน

When you read a sentence, you do not weigh every word equally. Attention gives a model that same power to focus on what matters. 🎯

Older models squeezed a whole sentence into one fixed vector. That bottleneck lost detail, especially for long inputs that carry many ideas.

Instead of one summary vector, attention lets the model look back at every input word whenever it needs to, picking what is relevant right now.

Attention assigns each word a weight between 0 and 1. Higher weight means more focus; the weights for one step always add up to 1.

The output is a weighted average of the input vectors. Words with big weights shape the result; words with tiny weights barely matter.

Translating "the cat sat" into French, the model can align each output word to the right source word instead of guessing from one blob.

Attention is soft: it spreads focus across all words by degree, rather than hard-picking just one. That makes it smooth and trainable.

To set the weights, the model scores how relevant each word is to the current step, then turns those scores into a probability spread.

Raw scores can be any number, so softmax squashes them into positive weights that sum to 1 and emphasize the largest score.

Because it can reach any word directly, attention keeps long-range links intact. The first word can still influence the last with no decay.

Attention freed models from reading strictly in order. That single idea became the foundation of the Transformer and modern language models. 🚀

Let us check the core intuition behind attention.

You learned that attention weights every input word by relevance and blends them into a focused output. That focus is what powers modern NLP. ✨