aboutsummaryrefslogtreecommitdiff
diff options
context:
space:
mode:
-rw-r--r--microposts/rnn-fsm.md2
1 files changed, 1 insertions, 1 deletions
diff --git a/microposts/rnn-fsm.md b/microposts/rnn-fsm.md
index a1e1315..032adc6 100644
--- a/microposts/rnn-fsm.md
+++ b/microposts/rnn-fsm.md
@@ -5,7 +5,7 @@ date: 2018-05-11
Related to [a previous micropost](#neural-turing-machine).
-[The slides from Toronto](http://www.cs.toronto.edu/~rgrosse/csc321/lec9.pdf) is a nice introduction to RNN (recurrent neural network) from a computational point of view. It states that RNN can simulate any FSM (finite state machine, a.k.a. finite automata abbr. FA) with a toy example computing the parity of a binary string.
+[These slides from Toronto](http://www.cs.toronto.edu/~rgrosse/csc321/lec9.pdf) is a nice introduction to RNN (recurrent neural network) from a computational point of view. It states that RNN can simulate any FSM (finite state machine, a.k.a. finite automata abbr. FA) with a toy example computing the parity of a binary string.
[Goodfellow et. al.'s book](http://www.deeplearningbook.org/contents/rnn.html) (see page 372 and 374) goes one step further, stating that RNN with a hidden-to-hidden layer can simulate Turing machines, and not only that, but also the *universal* Turing machine abbr. UTM (the book referenced [Siegelmann-Sontag](https://www.sciencedirect.com/science/article/pii/S0022000085710136)), a property not shared by the weaker network where the hidden-to-hidden layer is replaced by an output-to-hidden layer (page 376).