Learning HMM parameters Sushmita Roy BMI/CS Oct 21 st, 2014

Recall the three questions in HMMs How likely is an HMM to have generated a given sequence? – Forward algorithm What is the most likely “path” for generating a sequence of observations – Viterbi algorithm How can we learn an HMM from a set of sequences? – Forward-backward or Baum-Welch (an EM algorithm)

Learning HMMs from data Parameter estimation If we knew the state sequence it would be easy to estimate the parameters But we need to work with hidden state sequences Use “expected” counts of state transitions

Reviewing the notation States with emissions will be numbered from 1 to K – 0 begin state, N end state observed character at position t Observed sequence Hidden state sequence or path Transition probabilities Emission probabilities: Probability of emitting symbol b from state k

Learning without hidden information Learning is simple if we know the correct path for each sequence in our training set Estimate parameters by counting the number of times each parameter is used across the training set 5 C A G T begin end

Learning without hidden information Transition probabilities Emission probabilities Number of transitions from state k to state l Number of times c is emitted from k

Learning with hidden information 5 C A G T 0 begin end ???? if we don’t know the correct path for each sequence in our training set, consider all possible paths for the sequence estimate parameters through a procedure that counts the expected number of times each parameter is used across the training set

The Baum-Welch algorithm Also known as Forward-backward algorithm An Expectation Maximization (EM) algorithm EM is a family of algorithms for learning probabilistic models in problems that involve hidden information Expectation: Estimate the “expected” number of times there are transitions and emissions (using current values of parameters) Maximization: Estimate parameters given expected counts Hidden variables are the state transitions and emission counts

Learning parameters: the Baum-Welch algorithm algorithm sketch: – initialize parameters of model – iterate until convergence calculate the expected number of times each transition or emission is used adjust the parameters to maximize the likelihood of these expected values

The expectation step We need to know the probability of the symbol at t being produced by state k, given the entire sequence x Given these we can compute our expected counts for state transitions, character emissions We also need to know the probability of symbol at t and (t+1) being produced by state k, and l respectively given sequence x

Computing First we compute the probability of the entire observed sequence with the t th symbol being generated by state k Then our quantity of interest is computed as Obtained from the forward algorithm

To compute We need the forward and backward algorithm Forward algorithm f k (t) Backward algorithm b k (t) Computing

Using the forward and backward variables, this is computed as

The backward algorithm the backward algorithm gives us, the probability of observing the rest of x, given that we’re in state k after t characters A 0.4 C 0.1 G 0.2 T 0.3 A 0.2 C 0.3 G 0.3 T 0.2 begin end A 0.1 C 0.4 G 0.4 T 0.1 C A G T A 0.4 C 0.1 G 0.1 T 0.4

A 0.1 C 0.4 G 0.4 T 0.1 Example of computing A 0.4 C 0.1 G 0.2 T 0.3 A 0.2 C 0.3 G 0.3 T 0.2 begin end C A G T A 0.4 C 0.1 G 0.1 T 0.4

Steps of the backward algorithm Initialization ( t=T ) Recursion ( t=T-1 to 1 ) Termination Note, the same quantity can be obtained from the forward algorithm as well

Computing This is the probability of symbols at t and t+1 emitted from states k and l given the entire sequence x

Putting it all together Assume we are given J training instances x 1,..,x j,.. x J Expectation step – Using current parameter values compute for each x j Apply the forward and backward algorithms Compute – expected number of transitions between all pairs of states – expected number of emissions for all states Maximization step – Using current expected counts Compute the transition and emission probabilities

The expectation step: emission count We need the expected number of times c is emitted by state k x j : j th training sequences sum over positions where c occurs in x

The expectation step: transition count Expected number of times of transitions from k to l

The maximization step Estimate new emission parameters by: Just like in the simple case but typically we’ll do some “smoothing” (e.g. add pseudocounts) Estimate new transition parameters by

The Baum-Welch algorithm initialize the parameters of the HMM iterate until convergence – initialize, with pseudocounts – E-step: for each training set sequence j = 1…n calculate values for sequence j add the contribution of sequence j to, – M-step: update the HMM parameters using,

Baum-Welch algorithm example Given – The HMM with the parameters initialized as shown – Two training sequences TAG, ACG A 0.1 C 0.4 G 0.4 T 0.1 A 0.4 C 0.1 G 0.1 T 0.4 beginend we’ll work through one iteration of Baum-Welch

Baum-Welch example (cont) Determining the forward values for TAG Here we compute just the values that are needed for computing successive values. For example, no point in calculating f 2 (1) In a similar way, we also compute forward values for ACG

Baum-Welch example (cont) Determining the backward values for TAG Again, here we compute just the values that are needed In a similar way, we also compute backward values for ACG

Baum-Welch example (cont) determining the expected emission counts for state 1 contribution of TAG contribution of ACG *note that the forward/backward values in these two columns differ; in each column they are computed for the sequence associated with the column

Baum-Welch example (cont) Determining the expected transition counts for state 1 (not using pseudocounts) In a similar way, we also determine the expected emission/transition counts for state 2 Contribution of TAGContribution of ACG

Baum-Welch example (cont) Maximization step: determining probabilities for state 1

Computational complexity of HMM algorithms Given an HMM with S states and a sequence of length L, the complexity of the Forward, Backward and Viterbi algorithms is – this assumes that the states are densely interconnected Given M sequences of length L, the complexity of Baum-Welch on each iteration is

Baum-Welch convergence Some convergence criteria – likelihood of the training sequences changes little – fixed number of iterations reached Usually converges in a small number of iterations Will converge to a local maximum (in the likelihood of the data given the model)

Summary Three problems in HMMs Probability of an observed sequence – Forward algorithm Most likely path for an observed sequence – Viterbi – Can be used for segmentation of observed sequence Parameter estimation – Baum-Welch – The backward algorithm is used to compute a quantity needed to estimate the posterior of a state given the entire observed sequence