Memory is not a storehouse. It is a computation that never ends.
At every instant the network rewrites its hidden state by the rule
h_t = φ(W_xh x_t + W_hh h_{t-1} + b_h),
folding the newest observation into everything that has already been seen.
From that continuously updated vector it produces an output
y_t = g(W_ho h_t + b_o),
so that speech, prediction, or classification emerge as side-effects of an unbroken chain of transformations.
An elegant loop that turns time itself into the substrate of understanding.
显示更多