from d2l import tensorflow as d2l
import tensorflow as tfA recurrent neural network carries a hidden state \mathbf{h}_t across time steps, a learned summary of all input seen so far:
\mathbf{h}_t = \phi(\mathbf{W}_{xh}\mathbf{x}_t + \mathbf{W}_{hh}\mathbf{h}_{t-1} + \mathbf{b}).
Same weights at every step, so the parameter count is constant regardless of sequence length. Unbounded effective context (in principle), with no fixed-size window like an n-gram.
An RNN unrolled across three time steps; the same weights are reused at every step.
The naive form: two matrix multiplies, summed:
<tf.Tensor: shape=(3, 4), dtype=float32, numpy=
array([[-0.80866534, 0.42513677, -0.26260477, 0.42250693],
[-1.2269405 , -1.3283389 , -1.4689487 , -0.73261046],
[ 3.4852662 , 2.9564004 , 1.46526 , -1.4161625 ]],
dtype=float32)>
Equivalently, concatenate input and hidden and multiply by the concatenated weight matrix. Same result, one matmul:
<tf.Tensor: shape=(3, 4), dtype=float32, numpy=
array([[-0.8086653 , 0.42513683, -0.2626047 , 0.42250693],
[-1.2269405 , -1.3283387 , -1.4689488 , -0.7326105 ],
[ 3.4852662 , 2.9564002 , 1.46526 , -1.4161625 ]],
dtype=float32)>
The concatenate-then-multiply form is what most framework RNN implementations actually do.
Targets are the inputs shifted forward by one token; each step predicts the next token.
Train on gold prefixes; generate on the model’s own outputs. That mismatch is the rollout-error problem again.