62
4 Advanced Neural Networks
Transposed
convolution
Fig. 4.6 An intuitive description that can create checkerboard artifacts
4.2 Recurrent Neural Network and Backpropagation
So far, we have considered the following one-way propagation:
|h = T|x :=
m
|m l ( .
(4.10)
Then how can we treat, for example, time-series data
|x(t), t = 1, 2, 3, . . . , T ,
(4.11)
within neural network models? Putting this one by one into the neural network is
one way. However, as an example of (4.11), what if we have the following?
|x(t = 1) = |I
|x(t = 2) = |have
|x(t = 3) = |a
|x(t = 4) = |pen
|x(t = 5) = |.
(4.12)