The simulation shows a sequence of tokens moving along a glowing cell-state conveyor belt through forget, input, and output gates whose opening and closing is driven by a simple reactive function, visually mirroring how an LSTM decides what to discard, write, and expose at each timestep.
Pick a sequence type from the dropdown, use the speed slider to control how fast timesteps advance, and watch the stats panel report each gate's open/closed state; press Rebuild for a fresh random run or Pause to freeze the animation at any step.
Sequence type, playback speed, play/pause, rebuild
The 'Long Dependency' sequence mode reproduces the classic LSTM party trick: a strong signal injected near the start of the sequence stays visibly bright on the belt and is successfully carried all the way to a distant recall step, something a plain RNN's decaying gradients would make almost impossible to learn.
The simulation shows a sequence of tokens moving along a glowing cell-state conveyor belt through forget, input, and output gates whose opening and closing is driven by a simple reactive function, visually mirroring how an LSTM decides what to discard, write, and expose at each timestep.
The simulation shows a sequence of tokens moving along a glowing cell-state conveyor belt through forget, input, and output gates whose opening and closing is driven by a simple reactive function, visually mirroring how an LSTM decides what to discard, write, and expose at each timestep.
Pick a sequence type from the dropdown, use the speed slider to control how fast timesteps advance, and watch the stats panel report each gate's open/closed state; press Rebuild for a fresh random run or Pause to freeze the animation at any step.
The 'Long Dependency' sequence mode reproduces the classic LSTM party trick: a strong signal injected near the start of the sequence stays visibly bright on the belt and is successfully carried all the way to a distant recall step, something a plain RNN's decaying gradients would make almost impossible to learn.