Scribd
Upload
91 views

E9 205 - Machine Learning For Signal Processing

This document outlines the homework assignments for the course "E9 205 – Machine Learning for Signal Processing". It includes 4 problems: 1. Deriving the maximum likelihood solution for class prior probabilities in a generative classification model. 2. Using maximum likelihood to estimate parameters for a linear channel model, and extending it to a more complex Gaussian mixture model (GMM) of the source signal. 3. Implementing a GMM classifier on speech and music data, evaluating performance for different numbers of mixtures and covariance types. 4. Implementing an HMM classifier on the same data, evaluating the effect of number of states, mixtures, and covariance type.

Uploaded by

Uday Gulghane
Copyright
© © All Rights Reserved
Available Formats
Download as PDF, TXT or read online on Scribd
91 views

E9 205 - Machine Learning For Signal Processing

This document outlines the homework assignments for the course "E9 205 – Machine Learning for Signal Processing". It includes 4 problems: 1. Deriving the maximum likelihood solution for class prior probabilities in a generative classification model. 2. Using maximum likelihood to estimate parameters for a linear channel model, and extending it to a more complex Gaussian mixture model (GMM) of the source signal. 3. Implementing a GMM classifier on speech and music data, evaluating performance for different numbers of mixtures and covariance types. 4. Implementing an HMM classifier on the same data, evaluating the effect of number of states, mixtures, and covariance type.

Uploaded by

Uday Gulghane
Copyright
© © All Rights Reserved
Available Formats
Download as PDF, TXT or read online on Scribd
You are on page 1/ 2

E9 205 – Machine Learning for Signal Processing

Homework # 2
Due date: Sept. 26, 2018

(3:45 PM) for analytical and end of the day for coding part.
Coding assignment submitted to mlsp18 doT iisc aT gmail doT com

September 14, 2018

1. Maximum Likelihood Classification Consider a generative classification model with


K classes defined by prior probabilities p(Ck ) = πk and class-conditional densities p(φ|Ck )
where φ is the input feature vector. Suppose that a training data is given {φn , tn } for
n = 1, .., N and tn denotes a binary target vector of dimension K with components
tnj = δj,k if input pattern φn belongs to class k. Assuming that the data points are
drawn independently, show that the ML solution for prior probabilities is given by,
Nk
πk =
N
where Nk is the number of points belonging to class k. (Points 10)
2. Maximum Likelihood Linear Regression - Kiran is doing his PhD on acoustic channel
estimation. In order to estimate the channel characteristics, he designs an experiment in
which he records the ultra sound signal at the source as well as at the output of the
acoustic channel. Let xi , i = 1, ... , N and yi , i = 1, ... , N denote the feature sequence
corresponding to the source and channel outputs.
(a) He begins with an assumption of a linear model for the channel,

y = Ax + b + ǫ

where the source features x are assumed to be non-random and ǫ represents i.i.d.
channel noise ǫ ∼ N (0, σ 2 I). Given this model, his advisor Mohan recommends
the maximum
 likelihood (ML) method to estimate the parameters of the channel
A, b, σ . How will you solve the problem if you were Kiran ? (Points 10)
(b) While Kiran is successful in estimating the parameters of his model, Mohan is un-
happy with the results when the model is used to approximate a cell phone trans-
mission. Mohan proposes a more complex model where the source ultrasound data
xi , i = 1, ... , N is modeled as a Gaussian mixture model (GMM).
M
X
x∼ αm N (x, µm , Σm )
m=1

Further, the channel is modeled as a linear transformation of the GMM mean com-
ponents µ̂m = Aµm + b. The covariances are not affected in this model. With the
channel ouptuts yi , i = 1, ... , N , how will you help Kiran achieve his PhD faster by
solving for the channel parameters Am , bm , m = 1, ... , M assuming that the source
signal GMM is already estimated. Simplify your result. (Points 15)

3. Implementing GMM - A set of training and test examples of music and speech are
provided.
http : //www.leap.ee.iisc.ac.in/sriram/teaching/M LSP 18/assignments/speech music data.tar.gz

Using these examples,

a Generate spectrogram features - Use the log magnitude spectrogram as before with
a 64 component magnitude FFT (NFFT). In this case, the spectrogram will have
dimension 32 times the number of frames (using 25 ms with a shift of 10 ms).
b Train two GMM models with K-means initialization (for each class) separately each
with (i) 2 mixtures with diagonal covariance, (ii) 2 mixtures with full covariance and
(iii) 5-mixture components with diagonal/full covariance respectively on this data.
Plot the log-likelihood as a function of the EM iteration.
c Classify the test samples using the built classifiers and report the performance in
terms of error rate (percentage of mis-classified samples) on the text data.
e Discuss the impact on the performance for different number of mixture components,
diagonal versus full covariance ?

(Points 30)

4. Implementing HMM - Using the same data and features as above

a Write a code to implement the likelihood computation using the forward variable
after assuming a uniform flat start based initialization with 3 and 5 states per HMM
and GMM with 2 mixture components per state.
b Write a code to implement the Viterbi algorithm to decode the best state sequence
using the existing model.
c Use the Baum-Welch reestimation method to train HMMs with examples from music
and speech features.
d Classify the test examples and report the performance. How does the performance
change for different number of states per HMM, different number of mixture com-
ponents per GMM ? Will diagonal or full covariance GMM be a good choice ?

Hint - Use a flat-start to initialize the model and use scaling technique in implementation.
(Points 35)

You might also like