CUSTOMER_CODE
SMUDE
DIVISION_CODE
SMUDE
EVENT_CODE
OCTOBER15
ASSESSMENT_CODE BT9401_OCTOBER15
QUESTION_TYPE
DESCRIPTIVE_QUESTION
QUESTION_ID
37022
QUESTION_TEXT
Explain the implementation of Bayesian techniques with different steps.
SCHEME OF
EVALUATION
● Choose initial values for the hyperparameters  and . Suppose that n
samples x1, …., xn are drawn independently and identically distributed (i.i.d.)
according to the probability law p(x). Clearly, the probability that k of these n fall
in R is given by the binomial law.
● Train the network to minimize the total error function S(w) using a standard
algorithm for non-linear optimization.
● Repeat the estimation of values for  and  every few rounds of the
algorithm, with  calculated. This involves Hessian matrix evaluation and also
its eigenvalue spectrum.
● Repeat steps 1-3 for different random initial choices for the network
weights in order to find different local minima. In principle, a check should be
made that the different solutions are not simply related by a symmetry
transformation of the network.
● Repeat steps 1-4 for a selection of different network models, and compare
their evidences using (19). Those eigenvalues higher than the cutoff value are
only considered for evaluation of log determinant of the Hessian. If a committee
of networks is required to be used, it could be ideal to choose a selection of
better networks based on their evidences and then to use appropriate
techniques to calculate suitable weighting values.
QUESTION_TYPE
DESCRIPTIVE_QUESTION
QUESTION_ID
37024
QUESTION_TEXT
How to classify the corrupted inputs to obtain minimum error? Explain each
case.
SCHEME OF
EVALUATION
There are two analytically solvable cases of particular interest:
when some of the features are missing, and when they are
corrupted by a noise source with known properties. In each case,
our goal is to extract maximum possible information about the
underlying distribution as possible and use the Bayes decision
rule.
●
●
QUESTION_TYPE
Missing features
Noisy features
DESCRIPTIVE_QUESTION
QUESTION_ID
37026
QUESTION_TEXT
What you mean by Bagging and Boosting? Explain each.
SCHEME OF
EVALUATION
Bagging is our first confrontation with multi classifier systems,
where the output of the final classifier is based on the outputs of
various component classifiers. The decision rule which depends
on the vote among the component classifier represents an
elementary pooling method to integrate the outputs of component
classifiers.
Boosting is used to achieve better accuracy of the learning
algorithm. Here, as a first step, a classifier with a higher than
average accuracy is created. As the next step, new component
classifiers are added which then constitute a collection with a joint
decision rule having arbitrarily high accuracy on the training set.
This is what we call as “boosting” – which improves the
performance. To sum up this concept, this approach considers
each component classifier sequentially and trains it with that
subset of training data which provides the most useful information
for the current set of component classifiers. Classification of a test
point x is based on the outputs of the component classifiers.
QUESTION_TYPE
DESCRIPTIVE_QUESTION
QUESTION_ID
126066
QUESTION_TEXT
State and explain the nearest neighbor rule
SCHEME OF
EVALUATION
The nearest neighbor rule is a sub optimal procedure which leads to an
error rate greater than the minimum possible, the Bayes rate. With an
unlimited number of prototypes the error rate is never worse than twice
the Bayes rate. (2 marks)
The label θ1 associated with the nearest neighbor is a random variable
and the probability that θ1=wi is merely the posterior probability
P(wi/x1). When the number of samples is very large, it is reasonable to
assume that x1 is sufficiently close to x that P(wi/x1)≈ P(wi/x) (2 marks)
Because this is exactly the probability that nature will be in state wi, the
nearest neighbor rule is effectively matching probabilities with nature by
P(wm/x)≈max P(wi/x) then the Bayes decision rule always selects wm.
this rule allows us to partition the feature space into cells consisting of
all points closer to a given training points (2 marks)
All points in such a cell are thus labeled by the category of the training
point a so called Voronoi tessellation of the space. When P (wm/x) is
close to unity the nearest neighbor selection is almost always the same as
the Bayes selection. When P(wm/x)is close to 1э c, so that all classes are
essentially equally likely. (2 marks)
The selection made by nearest neighbor rule and Bayes decision rule are
rarely the same, but the probability of error is approximately 1-1 э c for
both. While more careful analysis is clearly necessary, these
observations should make the good performance of the nearest-neighbor
rule less surprising. (2 marks) Total 10 marks
QUESTION
DESCRIPTIVE_QUESTION
_TYPE
QUESTION
126067
_ID
QUESTION Write the general definition of pattern recognition and a note on elements
_TEXT
involved in recognizing a pattern input data.
a.
data
b.
sensor
c.
measurement
SCHEME d.
OF
EVALUAT e.
ION
f.
feature extraction
features
knowledge
g.
classification
h.
labels
scheme: (1*8=8+2=10 marks)
QUESTION_TYPE
DESCRIPTIVE_QUESTION
QUESTION_ID
126068
QUESTION_TEXT
Write a note on design cycle of pattern recognition.
a.
data collection
b.
choosing relevant features
SCHEME OF EVALUATION c.
choosing appropriate model
d.
training the classifiers
e.
evaluating the classifiers
(scheme:2*5=10marks)