Name: _______________________________________
Roll No:_______________________________________ 5. Which of the following can reduce some weights to
1. Which of the following functions are convex with exactly zero of a linear regression model, even in the
respect to weights (w and v)? non-extreme case:
a) wx+b a) Increasing input dimension
b) (wx+b - y)^2 b) Decreasing input dimension
c) - y log (1/(1+exp(-wx-b)) - (1-y) log (1 - 1/(1+exp(- c) L1 regularization (LASSO)
wx-b)) d) L2 regularization (Ridge, weight decay)
d) (v max(0,wx+b) + c - y )^2 6. Universal approximation theorem suggests that:
2. Which of the following is an ensemble? a) Any smooth function can be modeled to arbitrary
a) Neural network precision within a fixed interval by a neural
b) Support vector machine network, provided the number of hidden layers
c) Boosted trees with adaboost with non-polynomial activation can increase
d) Random forest arbitrarily
3. The training loss of which of the following is a convex b) Any non-polynomial function can be modeled to
function? arbitrary precision with a polynomial function,
a) Support vector machine with polynomial kernel provided the degree of the polynomial is allowed
b) Linear support vector regression to be as high as needed
c) Decision tree trained using weighted entropy c) Any smooth function can be modeled to arbitrary
minimization precision within a fixed interval by a single hidden
d) Neural network with at least one hidden layer layer neural network with non-polynomial
trained using binary cross entropy activation, provided the number of hidden units
4. What is the role of regularization in a machine learning can increase arbitrarily
model? d) Any function can be approximated in an interval by
a) To restrict the class of functions that can be linear basis functions, such as the Fourier series.
modeled.
b) To reduce the training error.
c) To introduce diversity in the inputs.
d) To reduce the gap between training and validation
errors.
Name: _______________________________________ 5. Which of the following can reduce some weights to
Roll No:_______________________________________ exactly zero of a linear regression model, even in the
1. Which of the following functions are convex with non-extreme case:
respect to weights (w and v)? a) Increasing input dimension
a) wx+b b) Decreasing input dimension
b) (wx+b - y)^2 c) L1 regularization (LASSO)
c) - y log (1/(1+exp(-wx-b)) - (1-y) log (1 - 1/(1+exp(- d) L2 regularization (Ridge, weight decay)
wx-b)) 6. Universal approximation theorem suggests that:
d) (v max(0,wx+b) + c - y )^2 a) Any smooth function can be modeled to arbitrary
2. Which of the following is an ensemble? precision within a fixed interval by a neural
a) Neural network network, provided the number of hidden layers
b) Support vector machine with non-polynomial activation can increase
c) Boosted trees with adaboost arbitrarily
d) Random forest b) Any non-polynomial function can be modeled to
3. The training loss of which of the following is a convex arbitrary precision with a polynomial function,
function? provided the degree of the polynomial is allowed
a) Support vector machine with polynomial kernel to be as high as needed
b) Linear support vector regression c) Any smooth function can be modeled to arbitrary
c) Decision tree trained using weighted entropy precision within a fixed interval by a single hidden
minimization layer neural network with non-polynomial
d) Neural network with at least one hidden layer activation, provided the number of hidden units
trained using binary cross entropy can increase arbitrarily
4. What is the role of regularization in a machine learning d) Any function can be approximated in an interval by
model? linear basis functions, such as the Fourier series.
a) To restrict the class of functions that can be
modeled.
b) To reduce the training error.
c) To introduce diversity in the inputs.
d) To reduce the gap between training and validation
errors.
7. Which of the following is a requirement for different 10. Which of the following is a sure sign of model
models to work well together as an ensemble? overfitting?
a) Each model should give better than a random a) The training loss is very low
chance output b) The training loss can be decreased further by
b) Each model should be independent of the others training the model longer
c) Models should be trained in cascade after c) Training loss is very low but the validation loss is
increasing the weight of the samples on which the high
previous models made errors d) Both training and validation losses are very low
d) Each model should be linear e) Both training and validation losses are high
8. Which of the following metrics for assessing binary
classification is unsuitable when there is severe class
imbalance:
a) Accuracy
b) F1-score
c) Precision
d) Recall
e) Area under receiver operating characteristic curve
(ROC or AU-ROC)
9. Which of the following is true for k-means clustering
algorithm:
a) If the algorithm converges fast, it means the k was
a good choice
b) There is no hyper-parameter for this algorithm
c) The algorithm is guaranteed to terminate in a finite
number of steps
d) It can discover clusters of arbitrary shapes
7. Which of the following is a requirement for different
models to work well together as an ensemble? 10. Which of the following is a sure sign of model overfitting?
a) Each model should give better than a random a) The training loss is very low
chance output b) The training loss can be decreased further by
b) Each model should be independent of the others training the model longer
c) Models should be trained in cascade after c) Training loss is very low but the validation loss is
increasing the weight of the samples on which the high
previous models made errors d) Both training and validation losses are very low
d) Each model should be linear e) Both training and validation losses are high
8. Which of the following metrics for assessing binary
classification is unsuitable when there is severe class
imbalance:
a) Accuracy
b) F1-score
c) Precision
d) Recall
e) Area under receiver operating characteristic curve
(ROC or AU-ROC)
9. Which of the following is true for k-means clustering
algorithm:
a) If the algorithm converges fast, it means the k was
a good choice
b) There is no hyper-parameter for this algorithm
c) The algorithm is guaranteed to terminate in a finite
number of steps
d) It can discover clusters of arbitrary shapes
