Machine Learning: Fall Exam
IEOR E4525 Spring 2020,
May 8th, 2020
1. Classiļ¬er Boundaries [20 pts]
Given the following classiļ¬ers:
1 A logistic regression classiļ¬er.
2 A logistic regression classiļ¬er trained on the input features x1 , x2 , x1(2), x2(2).
3 A nearest neighbors classiļ¬er.
4 A SVM with linear kernel.
5 A SVM with a polynomial kernel of degree d = 2.
K(x, x’) = (1 + xx’)2 (1)
6 A SVM with a Gaussian kernel
K(x, x0 ) = e
āγ|xāx’|2 (2)
with and intermediate (neither too large, not too small) value of the ā parameter.
7 A linear SVM classiļ¬er trained with the input features x1 , x2 , x1(2), x2(2).
8 A SVM with a Gaussian kernel and a very small value of ā .
9 A LDA classiļ¬er
10 A QDA classiļ¬er
11 A Classiļ¬cation Tree.
12 A Gaussian Naive Bayes classiļ¬er.
Indicate with classification boundary in Figure (1) was most likely gener–ated by each one of the classiļ¬ers above.
Note that more that one of the classiļ¬ers might have generated the same boundary.
If not indicated otherwise assume the classiļ¬ers were trained on the raw input features x = (x1, x2) ā R2.
Figure 1: Classiļ¬cation boundaries for Problem 1
2. Polynomial Kernel [20 points]
Given by x = (xA ; xB ) ā R2 and x’ = (xA(‘); xB(‘)) ā R2 describe the feature
mapping Φ( ) that generates the kernel:
K(x; x’) = (1 + āxT x’)3 = (1 + ā (xA xA(‘) + xB xB(‘)))3 (3)
3. Dense Neural Network [30 points]
We have a dense neural network with two hidden layers deļ¬ned by the following graph:
Figure 2: Network Architecture for Problem 5
where
⢠The two hidden layers have ReLU activation.
⢠the connection between the input an the ļ¬rst hidden layer is given by
and
⢠The connection between the ļ¬rst and second hidden layers is given by
and
⢠the connection between the second hidden layer and the output unit η is given by
W2 = (1 2 -1) (8)
and
b2 = (-3) (9)
⢠the last layer has logistic activation so that
⢠The network performance is assessed using the logistic loss
l(y, Ī·) = log (1 + eĪ· ) – yĪ· (11)
Given one input sample (x1 , x2 , x3 , x4 ) = (0, -2, 1, 1) and the networkās output label y = 1 compute:
(a) The ļ¬rst hidden layer activations (a1(1), a2(1)).
(b) The second hidden layer activations (a1(2), a2(2), a3(2)).
(c) The output Ė(y)(x)
With that we have
(d) Back propagate the errors Γα through the network layers
(e) Compute the gradients āW2/ā L and āb2/āL for the learnable parameters W2 , b2 of the output layer.
(f) Compute the gradient to the learnable parameters W1 , b1 of the second hidden layer
(g) Compute the gradient to the learnable parameters W0 , b0 of the ļ¬rst hidden layer
4. Tweedie Distribution: Generalized Linear Model [30 points]
The Tweedie distribution is frequently use to model the distribution of in- surance losses. This distribution belongs to an exponential family deļ¬ned by the probability density function
where the parameter μ > 0 is unknown, the parameter 1 < p < 2 is assumed to be known, and the function h(y; p) is the base measure that does not depend on μ .
(a) Write an expression for the sufficient statistic T (y) of the tweedie distribution:
(b) Write an expression for the canonical parameter η as a function of μ and p.
(c) Find the expression for the cumulant function A(Ī·)
(d) Derive an expression for the expected valueĖ(y) of the random variable y in terms of the parameters p and μ:
(e) Derive an expression for the variance of y in terms of the distribution parameters p and μ
(f) Assume we have training observations {xi , yi } for i = 1, . . . , N where we wish to assume that
⢠yi > 0 is an observed, positive valued labels.
⢠xi ā RD is our input data.
⢠p(yi jxi ) has a tweedie distribution with ļ¬xed, known parameter p but unknown parameter μ that depends on x.
⢠The canonical parameter ηi is linear in xi
Ī·(xi ) = xi(T)w + b (14)
Write the loss function L(w, b; {xi , yi } of the generalized tweedie re- gression model
(g) Write the gradient of the loss L with respect to w and b.
(h) What does the constrain μ > 0 implies for the allowed values of the canonical parameter? How would this afect the optimization problem of learning the parameters w ,b?. Explain brieļ¬y what you would do to resolve this issue.

![[SOLVED] IEOR E4525 Spring 2020 Machine Learning Fall Exam Java](https://assignmentchef.com/wp-content/uploads/2022/08/downloadzip.jpg)

![[SOLVED] POSIX Standards and Bindings](https://assignmentchef.com/wp-content/uploads/2022/08/downloadzip-1200x1200.jpg)
Reviews
There are no reviews yet.