RNN이 시퀀스를 읽는 방식

A recurrent neural network reads text like you do, taking one token, then the next, never the whole sentence in one gulp.

At each step the RNN keeps a small memory called the hidden state, a vector summarizing everything it has read so far.

The same cell runs again and again, once per word. That repeated, looping structure is exactly why we call it recurrent.

Each new word updates the hidden state, so context from earlier in the sentence quietly flows forward to later steps.

Conceptually the new state mixes the previous state with the current input through learned weights and an activation.

The RNN reuses the same weights at every position. That sharing lets it handle sentences of any length with one compact cell.

Each word first becomes a numeric embedding. The RNN never sees letters, only these dense vectors flowing in one by one.

Reading I love this, the state shifts after I, again after love, and the final state reflects the full short phrase.

For classification, the hidden state after the last word is often used as a summary of the entire sequence.

You can also read an output at every step, which is handy for tagging each word, like part-of-speech labels.

Because state carries the past, an RNN can finally notice that not flips the meaning of the word that follows it. 🔁

What does an RNN carry from one token to the next?

An RNN loops over tokens, updating a hidden state that remembers the past, so order and context finally influence its output. 🎯