1
$\begingroup$

The form of a Recurrent Neural Network (RNN) seems to resemble that of a hidden markov model. With a hidden markov model we have transitions between discrete states, as well as an emission variable that charts the observed measurement--from which we glean the hidden state. Hidden Markov models are usually fit using a forward/backward algorithm that is a specific instance of message passing based inference, as we would do in a graphical model. I guess if the model's hidden states are continuous, then we would use expectation propagation instead.

The point is, given that RNNs and hidden markov models look so similar, why can we use backpropagation through time for RNNs while we use message passing for hidden markov models. I was trying to understand the mathematical intuition behind this point.

Thanks.

$\endgroup$
1
  • 1
    $\begingroup$ Do this and this give you some elements ? $\endgroup$
    – TheCG
    Jul 21 '20 at 6:07
3
$\begingroup$

The two models have distinct levels of uncertainty. Once you have the parameters and inputs of an RNN, all hidden states of the RNN are uniquely determined. For an HMM, even if you know the parameters and emissions, you don't know the hidden states. The forward-backward algorithm is needed to integrate out the uncertainty over the hidden states of the HMM. You can see the difference in the likelihood function for an HMM compared to an RNN: the HMM likelihood has a sum over hidden states, while the RNN does not. When you take the gradient of the HMM likelihood, this sum remains, and you get the forward-backward algorithm.

If you were doing Bayesian inference in an RNN, then the hidden states would be uncertain (since the parameters are uncertain) and there would be a reason to do message passing/expectation propagation.

$\endgroup$
1
  • $\begingroup$ yes, this is great. This is exactly the type of answer I was looking for. Thanks for posting this. I get what you mean. I kinda see the RNN like a Kalman filter, in which case time moves forward and subsequent measurements in a time series are adjusted by the hidden state of the previous time step. For an HMM, it seems like states in the state space can be revisited over and over, so we have to integrate over the probability of being in each state times the probability of the measurement given that state. Yes, excellent response once again. $\endgroup$
    – krishnab
    May 24 at 19:33

Your Answer

By clicking “Post Your Answer”, you agree to our terms of service, privacy policy and cookie policy

Not the answer you're looking for? Browse other questions tagged or ask your own question.