3 ms·
Isn't this like an "auto-regressive feed forward neural net"[0]? > Instead of making predictions from a state that depends on the entire history, an autoregres
by tree_of_item 8y ago
Isn't this like an "auto-regressive feed forward neural net"[0]?
> Instead of making predictions from a state that depends on
the entire history, an autoregressive model directly predicts yt using only the k most recent inputs, xt−k+1,…,xt. This corresponds to a strong conditional independence assumption. In particular, a feed-forward model assumes the target only depends on the k most recent inputs. Google’s WaveNet[1] nicely illustrates this general principle.
[0] http://bair.berkeley.edu/blog/2018/08/06/recurrent/ http://bair.berkeley.edu/blog/2018/08/06/recurrent/
[1] https://arxiv.org/abs/1609.03499 https://arxiv.org/abs/1609.03499