Home▸Machine Learning & Neural Networks▸Bidirectional RNN: Forward-Backward Context Fusion

Bidirectional RNN: Forward-Backward Context Fusion

Watch two independent recurrent passes scan a sequence in opposite directions — a forward RNN reading left to right and a backward RNN reading right to left — and fuse into a single context vector at every position, exactly as a bidirectional RNN does.

Machine Learning & Neural Networks3DAdvanced60 FPS📱 Mobile-adapted

A bidirectional recurrent network is really two ordinary RNNs running independently over the same sequence — one reading left to right, one reading right to left — whose hidden states are concatenated at every position to build a context vector that knows both the past and the future. This simulator renders both chains as real recurrences (tanh(W·h + U·x + b), computed with genuine matrices, not a canned animation) advancing token by token in 3D, with a fusion bridge that lights up the moment both directions have reached a given position — visibly starting in the middle of the sequence and finishing last at the two ends. Toggle bidirectional mode off to see the same sequence processed by a plain forward-only RNN, where no future context ever reaches earlier tokens.

⚙ Under the hood

Watch two independent recurrent chains scan a sequence in opposite directions — forward left-to-right, backward right-to-left — and fuse into a single context vector at every position, exactly as a real bidirectional RNN computes it.

RNNbidirectionalLSTMsequence modelingdeep learninghidden state

3D · Three.js / WebGL renderer · 60 FPS target · runs fully client-side, no install