Lecture 12 | (1/5) Recurrent Neu
2019-10-28 本文已影响0人
Ysgc
https://www.youtube.com/watch?v=YYNNTrSROa4&list=PLp-0K3kfddPwz13VqV1PaMXF6V6dYdEsj&index=15&t=0s
(19fall)
Nonlinear autoregressive exogenous model
https://en.wikipedia.org/wiki/Nonlinear_autoregressive_exogenous_model
firstly proposed by M.I. Jordan
has infinite memory
黄框
红框
黄框
红框
X ---(W_0)---> Z_0 -----(activation)----> hidden ----(W_1)----> Z_1 -----(activation)----> Y <-----Div
Z_0(T-1)----(W_11)-----↘
X ---(W_0)---> Z_0 -----(activation)----> hidden ----(W_1)----> Z_1 -----(activation)----> Y <-----Div
the sign is += here!!!
from future the past can be predicted, just as the future can be predicted by the past
2 separate RNNs with different directions, and combine them before the output layer
