기억을 제어하는 게이트

A simple RNN struggles to remember things from far back. The LSTM was built to fix that by carefully managing what it keeps and forgets. 🧠

An LSTM carries a special cell state, a memory line that runs straight through every step so important signals survive long sequences.

The magic is in the gates: tiny learned switches that decide what to add, keep, or drop from memory at each time step.

Each gate uses a sigmoid to output values from 0 to 1, acting like a dimmer that lets some information through and blocks the rest.

The forget gate looks at the new input and decides how much of the old memory to erase. A 0 wipes it; a 1 keeps it fully.

The input gate controls how much fresh information gets written into the cell state, so only useful new signals are stored.

A candidate vector proposes what new content could enter memory. The input gate then scales how much of it actually gets added.

The new cell state blends old memory kept by the forget gate with fresh candidate values let in by the input gate.

The output gate decides what part of the cell state becomes the visible hidden state passed to the next step and the layer above.

Because gates keep the cell state mostly additive, gradients flow back many steps without vanishing. That is how an LSTM remembers long context.

So an LSTM has three gates guarding one memory line: forget, input, and output, each learned from data during training.

Test your grip on LSTM gates.

An LSTM protects a cell state with forget, input, and output gates. Those switches keep long-range memory alive where plain RNNs fail. ✅