Hidden Markov Model in TensorFlow

Jupyter Notebook:

Check out the Notebook for Examples.

Viterbi Algorithm

• Efficient way of finding the most likely state sequence.
• Method is general statistical framework of compound decision theory.
• Maximizes a posteriori probability recursively.
• Assumed to have a finite-state discrete-time Markov process.

Forward-Backward Algorithm

• The goal of the forward-backward algorithm is to find the conditional distribution over hidden states given the data.
• It is used to find the most likely state for any point in time.
• It cannot, however, be used to find the most likely sequence of states (see Viterbi)

Baum Welch Algorithm

• Expectation Maximization Inference of unknown parameters of a Hidden Markov Model.

Viterbi

Unrolled graph

Belief Propagation

Backtrack

Baum Welch and Forward-Backward

Forward-Backward

Re-estimate

Unrolled