Bidirectional RNN: Forward-Backward Context Fusion (2D)

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 2D-canvas companion renders both chains as real recurrences (tanh(W·h + U·x + b), computed with genuine matrices, not a canned animation) advancing token by token along a flat timeline, with per-position hidden-state vector bars and 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.