Font size
WorksheetsAI and Machine Learning Quiz
Total questions: 97
Worksheet time: 49mins
Which of the following are some aspects in which AI has transformed business?
Web searching and advertisement.
AI has not been able to transform businesses.
Eliminating the need for health care services.
Creating an AI-powered society.
Which of these are reasons for Deep Learning recently taking off? (Check the three options that apply.)
We have access to a lot more data.
We have access to a lot more computational power.
Deep learning has resulted in significant improvements in important applications such as online advertising, speech recognition, and image recognition.
Neural Networks are a brand new field.
Recall this diagram of iterating over different ML ideas. Which of the statements below are true? (Check all that apply.)
Being able to try out ideas quickly allows deep learning engineers to iterate more quickly.
Faster computation can help speed up how long a team takes to iterate to a good idea.
Recent progress in deep learning algorithms has allowed us to train good models faster (even without changing the CPU/GPU hardware).
It is faster to train on a big dataset than a small dataset.
Neural networks are good at figuring out functions relating an input x to an output y given enough examples. True/False?
True
False
Features of animals, such as weight, height, and color, are used for classification between cats, dogs, or others. This is an example of "structured" data, because they are represented as arrays in a computer. True/False?
True
False
A dataset is composed of age and weight data for several people. This dataset is an example of "structured" data because it is represented as an array in a computer. True/False?
True
False
Why is an RNN (Recurrent Neural Network) used for machine translation, say translating English to French? (Check all that apply.)
It is applicable when the input/output is a sequence (e.g., a sequence of words).
It can be trained as a supervised learning problem.
RNNs represent the recurrent process of Idea->Code->Experiment->Idea->....
It is strictly more powerful than a Convolutional Neural Network (CNN).
From the given diagram, we can deduce that Large NN models are always better than traditional learning algorithms. True/False?
False
True
Assuming the trends described in the previous question's figure are accurate (and hoping you got the axis labels right), which of the following are true? (Check all that apply.)
Increasing the size of a neural network generally does not hurt an algorithm's performance, and it may help significantly.
Increasing the training set size generally does not hurt an algorithm's performance, and it may help significantly.
Decreasing the training set size generally does not hurt an algorithm's performance, and it may help significantly.
Decreasing the size of a neural network generally does not hurt an algorithm's performance, and it may help significantly.
In logistic regression, which of the following best expresses what we want ŷ to tell us?
P(y = 1 | x)
σ(Wx)
P(y = ŷ | x)
σ(Wx + b)
Suppose that ŷ = 0.9 and y = 1. What is the value of the "Logistic Loss"? Choose the best option.
0.105
∞
0.005
Consider the Numpy array x = np.array([[[1], [2]], [[3], [4]]]). What is the shape of x?
(2, 2, 1)
(1, 2, 2)
(4,)
(2, 2)
Given arrays a = np.random.randn(3, 3) and b = np.random.randn(2, 1), what happens when computing c = a + b?
The computation cannot happen because it is not possible to broadcast more than one dimension.
c.shape = (2, 1)
c.shape = (2, 3, 3)
c.shape = (3, 3)
Given a = np.random.randn(1, 3) and b = np.random.randn(3, 3), what will be the shape of c = a * b?
c.shape = (3, 3)
The computation cannot happen because it is not possible to broadcast more than one dimension.
c.shape = (1, 3)
The computation cannot happen because the sizes don't match.
Suppose you have nx input features per example. What is the dimension of X?
(nx, m)
(m, 1)
(1, m)
(m, nx)
What is the result of np.dot(a, b) for a.shape = (12288, 150) and b.shape = (150, 45)?
c.shape = (12288, 45)
c.shape = (12288, 150)
c.shape = (150, 150)
The computation cannot happen because the sizes don't match.
Which of the following are true? (Check all that apply.)
a^[2] denotes the activation vector of the second layer.
w^[4]_3 is the column vector of parameters of the fourth layer and third neuron.
a^ denotes the activation vector of the second layer for the third example.
w^[4]_3 is the row vector of parameters of the fourth layer and third neuron.
a_3^[2] denotes the activation vector of the second layer for the third example.
The tanh activation is not always better than sigmoid activation function for hidden units because the mean of its output is closer to zero, and so it centers the data, making learning complex for the next layer. True/False?
True
False
Which of the following is a correct vectorized implementation of forward propagation for layer 2?
Z^[2] = W^[2] A^[1] + b^[2]; A^[2] = g^[2](Z^[2])
Z^[2] = W^[2] A^[1] + b^[2]; A^[2] = g(Z^[2])
Z^[1] = W^[1] X + b^[1]; A^[1] = g^[1](Z^[1])
Z^[2] = W^[2] X + b^[2]; A^[2] = g^[2](Z^[2])
You are building a binary classifier for recognizing cucumbers (y=1) vs. watermelons (y=0). Which one of these activation functions would you recommend using for the output layer?
sigmoid
ReLU
tanh
Leaky ReLU
Consider the following code: A = np.random.randn(4,3) B = np.sum(A, axis = 1, keepdims = True) What will be B.shape?
(4, 1)
(3,)
(1, 3)
(4,)
Suppose you have built a neural network. You decide to initialize the weights and biases to be zero. Which of the following statements is true?
Each neuron in the first hidden layer will perform the same computation. So even after multiple iterations of gradient descent, each neuron in the layer will be computing the same thing as other neurons.
Each neuron in the first hidden layer will perform the same computation in the first iteration. But after one iteration they will learn to compute different things because we have 'broken symmetry'.
The first hidden layer's neurons will perform different computations even in the first iteration.
Each neuron in the first hidden layer will compute the same thing, but neurons in different layers will compute different things, thus breaking symmetry.
A single output and single layer neural network that uses the sigmoid function as activation is equivalent to logistic regression. True/False?
True
False
You have built a network using the tanh activation for all the hidden units. You initialize the weights to relatively large values, using np.random.randn(...) * 1000. What will happen?
This will cause the inputs of the tanh to also be very large, thus causing gradients to be close to zero. The optimization algorithm will thus become slow.
This will cause the units to be 'highly activated' and thus speed up learning.
Gradient descent is not affected by whether the weights are large or small.
This will cause gradients to be large. You'll need a small alpha to prevent divergence, slowing learning.
We use the 'cache' in our implementation of forward and backward propagation to pass useful values to the next layer in the forward propagation. True/False?
False
True
Which of the following are 'parameters' of a neural network? (Check all that apply.)
b^[l] the bias vector.
W^[l] the weight matrices.
L the number of layers of the neural network.
g^[l] the activation functions.
Which of the following statements is true?
The deeper layers of a neural network are typically computing more complex features of the input than the earlier layers.
The earlier layers of a neural network are typically computing more complex features of the input than the deeper layers.
Vectorization allows us to compute a^[l] for all the examples on a batch at the same time without using a for loop. True/False?
True
False
If L is the number of layers of a neural network then dZ^[L] = A^[L] − Y. True/False?
True
False
There are certain functions with the following properties: (i) To compute the function using a shallow network circuit, you will need a large network; (ii) To compute it using a deep network circuit, you need only an exponentially smaller network. True/False?
True
False
What is the dimension of W^[l], the weight matrix associated with layer l?
W^[l] has shape (n^[l], n^[l−1])
W^[l] has shape (n^[l+1], n^[l])
W^[l] has shape (n^[l], n^[l+1])
W^[l] has shape (n^[l−1], n^[l])
If you have 10,000 examples, how would you split the train/dev/test set? Choose the best option.
60% train, 20% dev, 20% test.
98% train, 1% dev, 1% test.
33% train, 33% dev, 33% test.
In a personal experiment, an M.L. student decides to not use a test set, only train-dev sets. Which of the following is true?
He might be overfitting to the dev set.
He won't be able to measure the bias of the model.
Not having a test set is unacceptable under any circumstance.
He won't be able to measure the variance of the model.
If your Neural Network model seems to have high variance, what would be promising things to try? (Check all that apply.)
Add regularization.
Increase the number of units in each hidden layer.
Make the Neural Network deeper.
Get more test data.
You are building a classifier and your training error is 19% and dev error is 21%. What are promising things to try? (Check all that apply.)
Use a bigger network.
Get more training data.
Increase the regularization parameter lambda.
What is weight decay?
A regularization technique (such as L2 regularization) that results in gradient descent shrinking the weights on every iteration.
The process of gradually decreasing the learning rate during training.
A technique to avoid vanishing gradient.
Gradual corruption of weights in noisy data.
What happens when you increase the regularization hyperparameter lambda?
Weights are pushed toward becoming smaller (closer to 0).
Doubling lambda should roughly result in doubling the weights.
Gradient descent takes bigger steps.
Weights are pushed toward becoming bigger.
With the inverted dropout technique, at test time:
You do not apply dropout (do not randomly eliminate units) and do not keep the 1/keep_prob factor in the calculations used in training.
You apply dropout and keep the 1/keep_prob factor.
You do not apply dropout but keep the 1/keep_prob factor.
You apply dropout and do not keep the 1/keep_prob factor.
During training a deep neural network that uses tanh, gradients can become close to zero. What helps with this vanishing gradient problem?
Use Xavier initialization.
Increase the number of layers.
Use a larger regularization parameter.
Increase number of training cycles.
Which of the following actions increase the regularization of a model? (Check all that apply.)
Increase the value of the hyperparameter lambda.
Decrease the value of keep_prob in dropout.
Increase the value of keep_prob in dropout.
Decrease the value of the hyperparameter lambda.
Use Xavier initialization.
Why do we normalize the inputs x?
It makes the cost function faster to optimize.
It makes the parameter initialization faster.
It makes it easier to visualize the data.
Normalization is the same as regularization.
Which notation would you use to denote the 3rd layer's activations when the input is the 7th example from the 8th minibatch?
a^[3]{7}(8)
a^[8]{3}(7)
a^[8]{7}(3)
a^[3]{8}(7)
Which of these statements about mini-batch gradient descent do you agree with?
Training one epoch using mini-batch gradient descent is faster than training one epoch using batch gradient descent.
One iteration of mini-batch GD is faster than batch GD.
You should implement mini-batch GD without an explicit loop over batches.
Which of the following is true about batch gradient descent?
It is the same as mini-batch gradient descent when the mini-batch size is the same as the size of the training set.
It has as many mini-batches as examples.
It is the same as stochastic GD but without random elements.
Which of the following statements about Adam is False?
Adam should be used with batch gradient computations, not with mini-batches.
We usually use default values for β₁, β₂, ε.
The learning rate α in Adam usually needs tuning.
Adam combines the advantages of RMSProp and momentum.
Which of the following are true about hyperparameter search?
Choosing random values for the hyperparameters is convenient since we might not know in advance which hyperparameters are more important for the problem at hand.
When using random values for the hyperparameters they must be always uniformly distributed.
Choosing values in a grid is always better because it is more ordered.
When sampling from a grid, the number of values for each hyperparameter is larger than random.
Every hyperparameter, if set poorly, can have a huge negative impact on training, and so all hyperparameters are about equally important to tune well. True or False?
False
True
Even with enough computational power, is it better to babysit one model ('Panda' strategy)? True/False?
False
True
Knowing that the hyperparameter α should be in the range of 0.001 and 1.0, what is the recommended way to sample a value for α?
r = -3 * np.random.rand(); alpha = 10**r
r = 4 * np.random.rand(); alpha = 10**r
r = np.random.rand(); alpha = 0.001 + r * 0.999
r = -5 * np.random.rand(); alpha = 10**r
In batch normalization, if applied on the l-th layer of your neural network, what are you normalizing?
z^[l]
a^[l]
W^[l]
b^[l]
Which of the following are true about batch normalization?
One intuition behind why batch normalization works is that it helps reduce the internal covariance.
The parameter ε is used to accelerate convergence.
There is a global value of γ and β used for all hidden layers.
The parameters β and γ can't be trained using Adam or RMSProp.
After training a neural network with Batch Norm, how do you evaluate on a new test example?
Perform the needed normalizations using μ and σ² estimated using an exponentially weighted average across mini-batches.
Use the most recent mini-batch's μ and σ² to perform normalizations.
Duplicate the test example to simulate a full mini-batch.
Skip the normalization step.
Which of these statements about deep learning programming frameworks are true? (Check all that apply.)
A programming framework allows you to code deep learning with fewer lines than a lower-level language.
Good governance ensures open-source projects stay open and useful for all.
Deep learning programming frameworks require cloud-based machines to run.
The city asks for your help in defining the criteria for accuracy, runtime, and memory. How do you suggest they identify the criteria?
Suggest to them that they define which criterion is to be optimized. Then, set thresholds for the other two.
Suggest they focus on one criterion and eliminate the others.
Suggest they purchase more infrastructure to ensure the model runs quickly and accurately.
Why is it important to identify optimizing and satisficing metrics?
Identifying the metric types sets thresholds for satisficing metrics. This provides explicit evaluation criteria.
Knowing the metrics provides input for efficient project planning.
Identifying the optimizing metric informs the team which models to try first.
All metrics must be met for the model to be acceptable.
Why should you object to adding 1,000,000 citizens' images to the test set? (Check all that apply.)
This would cause the dev and test set distributions to become different.
The test set no longer reflects the distribution of data you care about.
The images do not have consistent x→y mapping.
A bigger test set will slow down training too much.
Human performance is <1%, training set error is 5.2%, dev error is 7.3%. What's the next step?
Train a bigger network to drive down the >4.0% training error.
Try an ensemble model to reduce bias and variance.
Get more data or regularization to reduce variance.
Validate the human data with a sample.
What best defines 'human-level performance' as a proxy for Bayes error?
The best performance of a specialist (ornithologist) or possibly a group of specialists.
The performance of a volunteer amateur.
The performance of the head of the City Council.
The average citizen of Peacetopia.
You observe a test error of 7.0% and dev error of 2.1%. What should you do? (Check all that apply.)
Try increasing regularization to reduce overfitting to the dev set.
Try decreasing regularization for better generalization with the dev set.
Get a bigger test set to increase its accuracy.
Increase the size of the dev set.
You achieve: Human-level performance 0.10%, Training set error 0.05%, Dev set error 0.05%. Which of the following are likely? (Check all that apply.)
The model has recognized emergent features humans cannot.
Pushing to even higher accuracy will be slow due to difficulty identifying bias.
There is still avoidable bias.
This result is not possible.
You are building a model to detect road signs and traffic signals. Multiple labels can appear in an image. Which activation function is most appropriate for the output layer?
Sigmoid
Softmax
Linear
ReLU
You are doing error analysis. Which dataset should you manually review to understand what the algorithm got wrong?
500 images on which the algorithm made a mistake
500 randomly chosen images
10,000 images on which the algorithm made a mistake
10,000 randomly chosen images
The dev and test sets should come from which data distribution?
The front-facing camera (the data you care about most)
The same distribution as the training set
A combination of training set and online images
The internet, for variety
After fixing mislabeled data in the dev set, which statements are true? (Check all that apply.)
You should also correct the incorrectly labeled data in the test set, so dev and test continue to come from the same distribution.
You don't necessarily need to fix mislabeled data in the training set, as long as the dev and test sets match.
You should never correct data in the test set.
You must correct training data to keep it aligned with dev/test distributions.
To recognize red and green lights, you've been using Approach A: an end-to-end model from image to output. A teammate suggests a two-step approach (Approach B). Which is more end-to-end?
False (Approach B is not more end-to-end)
True
End-to-end approaches don't require hand-designed features, only a large enough model. True/False?
True
False
What do you think applying this filter to a grayscale image will do?
Detect horizontal edges.
Detect vertical edges.
Detecting image contrast.
Detect 45-degree edges.
Your input is a 300x300 RGB image. A fully connected layer has 100 neurons. How many parameters does it have?
27,000,100
9,000,001
9,000,100
27,000,001
Your input is 256x256 RGB. You use 128 filters of size 7x7. How many parameters are there?
18944
18816
6400
1233125504
An input volume 127x127x16 is convolved with 32 filters of 5x5, stride=2, no padding. What is the output volume?
62 x 62 x 32
62 x 62 x 16
123 x 123 x 32
123 x 123 x 16
You have an input of 31x31x32. With padding=1, what is the new dimension?
33x33x32
32x32x32
33x33x33
31x31x34
You apply SAME convolution with padding=?. Input is 64x64x32, 40 filters of 9x9, stride=1.
4
0
6
8
Input is 66x66x21, apply max pooling with stride=3 and filter=3. Output volume?
22 x 22 x 21
22 x 22 x 7
66 x 66 x 7
21 x 21 x 21
Which of the following are benefits of convolutional layers? (Check all that apply.)
Convolutional layers are good at capturing translation invariance.
It reduces the total number of parameters, thus reducing overfitting through parameter sharing.
It reduces the computations in backpropagation since we omit the convolutional layers.
The sparsity of connections and weight sharing allow ConvNets to be trained with smaller datasets. True/False?
True
False
Which of the following are typical in ConvNets? (Check all that apply.)
FC layers in the last few layers
Multiple CONV layers followed by a POOL layer
Multiple POOL layers followed by a CONV layer
FC layers in the first few layers
In LeNet-5, as we go deeper in the network, the number of channels increases, and spatial dimensions decrease. True/False?
True
False
Which of the following equations captures computations in a ResNet block?
a^[l+2] = g(W^[l+2]g(W^[l+1]a^[l] + b^[l+1]) + b^[l+2]) + a^[l]
a^[l+2] = g(W^[l+2]g(W^[l+1]a^[l] + b^[l+1]) + b^[l+2])
a^[l+2] = g(W^[l+2]a^[l] + b^[l+2]) + a^[l+1]
a^[l+2] = g(W^[l+2]g(W^[l+1]a^[l] + b^[l+1]) + a^[l+1]) + a^[l]
Which are true about Residual Networks? (Check all that apply.)
Using a skip-connection helps the gradient to backpropagate and thus helps you to train deeper networks.
The skip-connections compute a complex non-linear function of the input.
A ResNet with L layers has ~L² skip connections.
Skip-connections make it easy for the network to learn a complex mapping.
Which of the following about 1x1 convolution are true? (Check all that apply.)
You can use a 1x1 convolutional layer to reduce n_C but not n_H and n_W.
You can use a 2D pooling layer to reduce n_H and n_W, but not n_C.
You can use a 1x1 convolutional layer to reduce n_H, n_W, and n_C.
You can use a 2D pooling layer to reduce n_H, n_W, and n_C.
Which are true about Inception Networks? (Check all that apply.)
A single inception block allows the use of 1x1, 3x3, 5x5 convolutions and pooling.
Inception blocks use 1x1 convolutions to reduce input size before 3x3 and 5x5 convolutions.
Inception networks incorporate a variety of architectures and regularization similar to dropout.
Making an inception network deeper always improves performance.
Why use open-source implementations of ConvNets? (Check all that apply.)
Parameters trained for one CV task are often useful as pre-training for others.
It's a convenient way to start using complex architectures.
You can use one CV model to augment data for another task.
Competition tricks (like cropping) are widely used in production systems.
In Depthwise Separable Convolution you: (Check all that apply.)
Perform two steps of convolution.
The final output is of the dimension n_out x n_out x n_c'
Perform one step of convolution.
Each filter convolves with all input channels (Depthwise)
Each Depthwise filter convolves with one input channel.
Suppose your training examples are sentences (sequences of words). Which of the following refers to the sᵗʰ word in the rᵗʰ training example?
x⁽ʳ⁾<s>
x<ˢ>(ʳ)
x<ʳ>(ˢ)
x(ˢ)<ʳ>
A Transformer Network processes sentences from left to right, one word at a time.
False
True
The major innovation of the transformer architecture is combining the use of LSTMs and RNN sequential processing.
False
True
What letter does the "?" represent in the following representation of Attention? Attention(Q, K, V) = softmax(QKᵀ / √dₖ)V
k
v
q
t
Which of the following statements represents Key (K) as used in the self-attention calculation?
K = qualities of words given a Q
K = the order of the words in a sentence
K = interesting questions about the words in a sentence
K = specific representations of words given a Q
Which of these is a good criterion for a good positional encoding algorithm?
Distance between any two time-steps should be consistent for all sentence lengths.
The algorithm should be able to generalize to longer sentences.
It should output a unique encoding for each time-step (word's position in a sentence).
None of these.
Suppose you learn a word embedding for a vocabulary of 10000 words. Then the embedding vectors could be 10000 dimensional, so as to capture the full range of variation and meaning in those words.
False
True
True/False: t-SNE is a non-linear dimensionality reduction technique.
True
False
Which of these equations do you think should hold for a good word embedding? (Check all that apply)
e₍boy₎ − e₍girl₎ ≈ e₍brother₎ − e₍sister₎
e₍boy₎ − e₍brother₎ ≈ e₍girl₎ − e₍sister₎
e₍boy₎ − e₍girl₎ ≈ e₍sister₎ − e₍brother₎
e₍boy₎ − e₍brother₎ ≈ e₍sister₎ − e₍girl₎
True/False: The most computationally efficient formula for Python to get the embedding of word 1021, if C is an embedding matrix and o₁₀₂₁ is a one-hot vector for word 1021, is Cᵀ * o₁₀₂₁.
False
True
Suppose you have a 10000 word vocabulary, and are learning 500-dimensional word embeddings. The word2vec model uses the softmax function. Which of these statements are correct? (Check all that apply)
θₜ and e𝚌 are both 500 dimensional vectors.
θₜ and e𝚌 are both trained with an optimization algorithm such as Adam or gradient descent.
After training, we should expect θₜ to be very close to e𝚌 when t and c are the same word.
What will be the shape of c = a * b, if a = np.random.randn(3, 3) and b = np.random.randn(3, 1)?
This will invoke broadcasting, so b is copied three times to become (3, 3), and * is an element-wise product so c.shape will be (3, 3)
This will multiply a 3x3 matrix with a 3x1 vector, thus resulting in a 3x1 vector.
It will lead to an error since you cannot use "*" to operate on these two matrices.
This will invoke broadcasting and then matrix multiplication, resulting in c.shape = (3, 3)
Suppose you have a 10000 word vocabulary, and are learning 500-dimensional word embeddings. The GloVe model minimizes this objective:
min ∑ᵢ₌₁¹⁰⁰⁰⁰ ∑ⱼ₌₁¹⁰⁰⁰⁰ f(Xᵢⱼ)(θᵢᵀ eⱼ + bᵢ + bⱼ' − log Xᵢⱼ)²
True/False: Xᵢⱼ is the number of times word j appears in the context of word i.
False
You have trained word embeddings using a text dataset of m₁ words. You are considering using these word embeddings for a language task, for which you have a separate labeled dataset of m₂ words.
Keeping in mind that using word embeddings is a form of transfer learning, under which of these circumstances would you expect the word embeddings to be helpful?
m₁ ≫ m₂
m₁ ≪ m₂
