Download presentation

Presentation is loading. Please wait.

Published byGeoffrey Bradley Modified over 4 years ago

1
Hidden Markov Models Reading: Russell and Norvig, Chapter 15, Sections 15.1-15.3

2
Recall for Bayesian networks: General question: Given query variable X and observed evidence variable values e, what is P(X|e)?

3
Dynamic Bayesian Networks General dynamic Bayesian network: any number of random variables, which can be discrete or continuous Observations are taken in time steps. At each time step, observe some of the variables (evidence variables). Other variables are unobserved or “hidden”.

4
Simple example of HMM (adapted from Russell and Norvig, Chapter 15) You are a graduate student in a windowless office with no phone and your network connection is down. The only way you can get information about the weather outside is whether or not your advisor shows up carrying an umbrella.

5
HMM for this scenario: Evidence variable: Umbrella {T, F} Hidden variable Rain {T, F} Markov model since R t depends only on R t-1. Rain t-1 Umbrella t-1 Umbrella t Rain t Umbrella t+1 Rain t+1 Rain t-1 P(Rain t ) T0.7 F0.3 Rain t P(Umbrella t ) T0.9 F0.2 P(Rain 0 =T ) = 0.5

6
Inference in Hidden Markov Models Inference tasks: – Filtering (or monitoring): Computing belief state―posterior distribution over current state, given all evidence to date: “Given that my advisor has had an umbrella for the last three days, what’s the probability it is raining today?”

7
Inference in Hidden Markov Models – Prediction: Computing posterior distribution over the future state, given all evidence to date: “Given that my advisor has had an umbrella for the last three days, what’s the probability it will rain the day after tomorrow?”

8
– Smoothing (or hindsight): Computing posterior probability over a past state, given all evidence up to the present: “Given that my advisor has had an umbrella for the last three days, what’s the probability it rained yesterday?

9
– Most likely explanation: Given a sequence of observations, finding the sequence of states most likely to have generated those observations: “Given that my advisor has had an umbrella for the last three days, what’s the most likely sequence of weather over the past 3 days?”

10
Inference algorithms Filtering: Can use recursive estimation

11
Inference algorithms Filtering: Can use recursive estimation The value of the first term,, is given explicitly in the network.

12
Inference algorithms Filtering: Can use recursive estimation The value of the first term,, is given explicitly in the network. The value of the second term is:

13
Inference algorithms Filtering: Can use recursive estimation The value of the first term,, is given explicitly in the network. The value of the second term is: Thus:

14
Inference algorithms From the network, we have everything except P(x t, e 1:t ). Can estimate recursively. Thus:

15
Umbrella example: Day 1: Umbrella 1 = U 1 = T Prediction t = 0 to t = 1: Updating with evidence for t=1:

16
Prediction t = 1 to t = 2: Updating with evidence for t=2: Why does probability of rain increase from day 1 to day 2?

17
Hidden Markov Models: Matrix Representations Transition model: P(X t | X t 1 ) = T (S S matrix) where For umbrella model: Sensor model: P(e t | X t = i ) =O (S S diagonal matrix) where For umbrella model:

18
Speech Recognition Task: Identify sequence of words uttered by speaker, given acoustic signal. Uncertainty introduced by noise, speaker error, variation in pronunciation, homonyms, etc. Thus speech recognition is viewed as problem of probabilistic inference.

19
Speech recognition typically makes three assumptions: 1.Process underlying change is itself “stationary” i.e., state transition probabilities don’t change 2.Current state X depends on only a finite history of previous states (“Markov assumption”). –Markov process of order n: Current state depends only on n previous states. 3.Values e t of evidence variables depend only on current state X t. (“Sensor model”)

20
Speech Recognition Input: acoustic signal Inference: P(words | signal) Bayes rule: P(words | signal) = P(signal | words) P(words) P(signal | words): acoustic model – pronunciation model (for each word, distribution over possible phone sequences) – signal model (distribution of features of acoustic signal over phones) P(words): language model – prior probability of each utterance (e.g., bigram model)

21
Russell and Norvig, Artificial Intelligence: A Modern Approach, Chapter 15

23
Phone model P( phone | frame features) = P(frame features| phone) P(phone) P(frame features| phone) often represented by Gaussian mixture model

24
Pronunciation model Now we want P (words|phones 1:t ) = P(phones 1:t | words) P(words) Represent P(phones 1:t | words) as an HMM

27
More Generally: Components of an HMM Rain t-1 Umbrella t-1 Umbrella t Rain t Umbrella t+1 Rain t+1 Rain t-1 P(Rain t ) T0.7 F0.3 Rain t P(Umbrella t ) T0.9 F0.2 P(Rain 0 =T ) = 0.5 Model consists of sequence of hidden states, sequence of observation states, probability of each hidden state given previous hidden state, probability of each hidden state given current observation, and prior probability of first hidden state.

28
Rain t-1 Umbrella t-1 Umbrella t Rain t Umbrella t+1 Rain t+1 Rain t-1 P(Rain t ) T0.7 F0.3 Rain t P(Umbrella t ) T0.9 F0.2 P(Rain 0 =T ) = 0.5 Model consists of sequence of hidden states, sequence of observation states, probability of each hidden state given previous hidden state, probability of each hidden state given current observation, and prior probability of first hidden state. Possible states: S = {S 1,..., S N }

29
Rain t-1 Umbrella t-1 Umbrella t Rain t Umbrella t+1 Rain t+1 Rain t-1 P(Rain t ) T0.7 F0.3 Rain t P(Umbrella t ) T0.9 F0.2 P(Rain 0 =T ) = 0.5 Model consists of sequence of hidden states, sequence of observation states, probability of each hidden state given previous hidden state, probability of each hidden state given current observation, and prior probability of first hidden state. State transition probabilities A = [a ij ], a ij =P(q t +1=S j |q t =S i )

30
Rain t-1 Umbrella t-1 Umbrella t Rain t Umbrella t+1 Rain t+1 Rain t-1 P(Rain t ) T0.7 F0.3 Rain t P(Umbrella t ) T0.9 F0.2 P(Rain 0 =T ) = 0.5 Model consists of sequence of hidden states, sequence of observation states, probability of each hidden state given previous hidden state, probability of each hidden state given current observation, and prior probability of first hidden state. Possible observations (or “emissions”): V={v 1,..., v M }

31
Rain t-1 Umbrella t-1 Umbrella t Rain t Umbrella t+1 Rain t+1 Rain t-1 P(Rain t ) T0.7 F0.3 Rain t P(Umbrella t ) T0.9 F0.2 P(Rain 0 =T ) = 0.5 Model consists of sequence of hidden states, sequence of observation states, probability of each hidden state given previous hidden state, probability of each hidden state given current observation, and prior probability of first hidden state. Observation (emission) probabilities: B = [b j (m)] b j (m)=P(O t =v m |q t =S i )

32
Rain t-1 Umbrella t-1 Umbrella t Rain t Umbrella t+1 Rain t+1 Rain t-1 P(Rain t ) T0.7 F0.3 Rain t P(Umbrella t ) T0.9 F0.2 P(Rain 0 =T ) = 0.5 Model consists of sequence of hidden states, sequence of observation states, probability of each hidden state given previous hidden state, probability of each hidden state given current observation, and prior probability of first hidden state. Initial state probabilities: = [ i ] i =P(q 1 =S i )

33
Learning an HMM Baum-Welch algorithm (also known as “forward-backward algorithm), similar to Expectation-Maximization

Similar presentations

© 2019 SlidePlayer.com Inc.

All rights reserved.

To make this website work, we log user data and share it with processors. To use this website, you must agree to our Privacy Policy, including cookie policy.

Ads by Google