代写APS 360 Applied Fundamentals of Machine Learning Midterm Test Winter 2022调试数据库编程

APS 360

Winter 2022          Midterm Test

Applied Fundamentals of Machine Learning

Date :    March 04, 2022

Question 1.   [10 marks]

Circle the best answer for each of the questions below. Do not circle more than one answer per question.

Part (a)   [1 mark]

Which of these problems is best described as an unsupervised learning problem?

(A) Generating a random pictures

(B) Representation learning with a contrastive loss * (1 marks)

(C) Traversing a neural network graph during a forward pass

(D) Predict stock prices from the size of a company, given a dataset of companies size and prices

Part (b)   [1 mark]

Which of these best describes the concept of inductive  bias?

(A) Bias of a machine learning algorithm learned from a training data set

(B) The assumptions we incorporate in our model to help learn to solve a problem * (1 marks)

(C) Learning from an unbalanced data set

(D) A model with high precision but low recall

Part (c)   [1 mark]

Which loss would be most suitable  for training a neural network for regression?

(A) Mean squared error plus weight decay * (1 marks)

(B) Backpropagation

(C) Cross entropy

(D) Delta rule

Part (d)   [1 mark]

Which of the following activation func-tions are not compatible with back- propagation?

(A) f(x) = tanh(x)

(B) f(x) = max(0, x)

(C) f(x) = sign(x) * (1 marks)

(D) f(x) = (1 + e −x ) −1

Part (e)   [1 mark]

Which of the following is true about con- volutional neural networks?

(A) Pooling/stride reduce the number of channels of an input image/feature map.

(B) Convolution is used to reduce the height/width of an input image/feature map.

(C) Convolving an input image/feature map having any number of channels with a single kernel produces a fea-ture map with one channel. * (1 marks)

(D) The order of operations used in a convolutional neural network is typically “Convolution-Pooling-Activation”.

Part (f)   [1 mark]

Assume an input image of size 64×64 with 3 channels, and a model with a single con- volutional  layer  having  2  kernels  of  size 3x3, and 3 kernels of size 5×5.  What are the total number of convolutional weights you need to learn?   Assume there are no bias terms.

(A) (2 × 3 × 3 × 3) + (3 × 3 × 5 × 5) * (1 marks)

(B) (2 × 3 × 3) + (3 × 5 × 5)

(C) (64 × 64 × 3) + (2 × 3 × 3) + (3 × 5 × 5)

(D) 64 × 64 × 3 × 3 × 5 × 5

Part (g)   [1 mark]

Which of these computes the gradients in PyTorch?

(A) optimizer.zero_grad()

(B) optimizer.step()

(C) loss.backward() * (1 marks)

(D) loss = criterion(model(inputs), labels)

Part (h)   [1 mark]

Which of the following is correct?

(A) Smaller batch sizes typically allow using larger learning rates.

(B) We optimize the weights on training set and the hyper-parameters on the test set.

(C) We can increase the learning rate as training progresses.

(D) For the problems discussed in class, neural networks are typically unable to find the global optimal. 1 (

(E) marks)Using a standard adaptive learning rate method reduce the memory usage.

Part (i)   [1 mark]

Suppose we have an input image of size 3×15×15  and  we  apply  a  kernel  of  size 3×4×4 with padding of size 2 and stride of 1, followed by a pooling with kernel size of 4×4 and strides of 4.  What are the di- mensions of the output?

(A) 3×4×4

(B) 1×4×4 * (1 marks)

(C) 3×8×8

(D) 1×8×8

Part (j)   [1 mark]

Which of the following is incorrect about convolutional neural networks?

(A) They are more efficient compared to fully-connected net-works.

(B) As we go deeper, height and width are decreased while depth is increased.

(C) We can use either convolution with strides or pooling layers to reduce the depth. * (1 marks)

(D) Kernels of size 3×3 can approximate bigger kernel sizes.

Question 2.   [10 marks]

In this question we will consider neural networks (NNs) with sigmoid activation functions of the form.

If we denote by vj(l)  the value of neuron j at layer l its value is computed as

 

where L is the number of neurons in the previous layer (i.e. layer l − 1).

In the following questions you will have to design NNs that compute functions of two Boolean inputs X1 and X2 . Recall that Boolean variable can only take on of two value:  0 or 1.  Given that the outputs of the sigmoid units are real numbers Y in the interval (0.0, 1.0), we will treat the final output as Boolean by considering it as 1 if greater than 0.5 and 0 otherwise.

Feel free to use a calculator or your computer for this question. Part (a)   [4 marks]

Determine three weights ω0 ,ω 1 ,ω2 , for a single layer NN with two inputs X1  and X2  that implements the logical OR function Y  = X1  OR X2 .   Note that, as per the equation above, ω0  is the bias weight (sometimes called b). The definition of X1  OR X2  is provided in the table below.

Part (b)   [4 marks]

Determine three weights ω0 ,ω 1 ,ω2 , for a single layer NN with two inputs X1  and X2  that implements the logical AND function Y = X1  AND X2 . The definition of X1  AND X2  is provided in the table below.

Part (c)   [2 marks]

Is it possible to implement the XOR logical function with a single layer NN? The definition of X1  XOR X2 is provided in the table below. Explain your answer.

Question 3.   [10 marks]

In the following network, we use softmax for the output activation function.

Part (a)   [2 marks]

In the figure above, what is the missing output value? Why?

Part (b)   [2 marks]

Assuming we are using one-hot encoding and the set of labels is {0,1,2,3}, What is the predicted label? Why?

Part (c)   [3 marks]

Assuming we have  1000 samples in a training set.   How  many times do we update the weights of the network in an epoch in the case of:

- i) Stochastic Gradient Descent?

- ii) Mini-batch gradient descent with batch size of 50?

- iii) Batch gradient Descent?

Part (d)   [3 marks]

Consider the following binary classification problem.

Using the following linear classifier:

f(X) = sign(ω0 + ω1x1 + ω2x2 ),

where sign(z) returns +1 (class 1, diamonds) for z > 0, otherwise it returns -1 (class 2, circles).

We try to minimize the following loss function in w = (ω0 ,ω 1 ,ω2 ) space:

J(w) = L(w) + λω0(2) ,

where L(w) = N/1 (total number of misclassif ied samples), N is the total number of samples and λ = 109 is a large constant. Which of the following is a possible decision boundary after the training? Explain your choice.

Question 4.   [12 marks]

class  SoDeep(nn.Module): 

def      init     (self):

super(SoDeep,  self) .     init     ()

self.layer1  =  nn.Linear(in_features=32*32,  out_features=64) 

self.layer2  =  nn.Linear(in_features=64,  out_features=128)

self.layer3  =  nn.Linear(in_features=128,  out_features=4)

def  forward(self,  img):

flattened  =  img.view(-1,  32*32)

x  =  self.layer1(flattened)

x  =  self.layer2(x)

return  self.layer3(x)

You are given a neural network architecture written in pytorch, as above.  Note:  You do not have to do the arithmetic to calculate any of the answers, writing the expression to calculate it is sufficient.

Part (a)   [2 marks]

How many weights does layer1 have?

Part (b)   [2 marks]

How many neurons does layer2 have?

Part (c)   [2 marks]

How many hidden layers are there in this model?