Hands-on AI II

Question 1

Which of the following are true about the ‘softmax’ function?
a. It is often used in the output layer of a multi-class classification network.
b. It converts a vector of real numbers (logits) into a probability distribution, where all elements are between 0 and 1 and sum to 1.
c. It is a generalization of the sigmoid function to multiple classes.
d. It is defined as σ(z)ᵢ = zᵢ / Σzⱼ.

Answer 1

a. It is often used in the output layer of a multi-class classification network.
b. It converts a vector of real numbers (logits) into a probability distribution, where all elements are between 0 and 1 and sum to 1.
c. It is a generalization of the sigmoid function to multiple classes.

Question 2

*Which of the following statements are correct regarding tabular data?
a. Each row typically represents a sample.
b. Each column typically represents a feature.
c. It can be used in both supervised and unsupervised learning.
d. It has a fixed set of properties (columns) for every record (row).

Answer 2

a. Each row typically represents a sample.
b. Each column typically represents a feature.
c. It can be used in both supervised and unsupervised learning.
d. It has a fixed set of properties (columns) for every record (row).

Question 3

How does a one-to-many RNN architecture typically work?
a. It takes a single input vector and generates a sequence of output vectors.
b. An example application is image captioning, where the input is a feature vector from an image and the output is a sentence.
c. It processes a sequence of inputs to produce a single output.
d. It is the standard architecture for sentiment classification.

Answer 3

a. It takes a single input vector and generates a sequence of output vectors.
b. An example application is image captioning, where the input is a feature vector from an image and the output is a sentence.

Question 4

What is ‘perplexity’ in the context of language modeling?
a. It is a measure of how well a probability model predicts a sample.
b. It is calculated as the exponentiation of the cross-entropy loss.
c. A lower perplexity indicates a better language model.
d. A perplexity of N means that the model is as confused as if it had to choose uniformly among N choices at each step.

Answer 4

a. It is a measure of how well a probability model predicts a sample.
b. It is calculated as the exponentiation of the cross-entropy loss.
c. A lower perplexity indicates a better language model.
d. A perplexity of N means that the model is as confused as if it had to choose uniformly among N choices at each step.

Question 5

**Which of the following statements are true about SMILES (Simplified Molecular Input Line Entry Specification)?
a. It is a string sequence representation of a molecule.
b. A molecule can have different SMILES representations.
c. A molecular graph can be transformed into a SMILES string and the other way around.
d. A molecule can only have exactly one SMILES representation.

Answer 5

a. It is a string sequence representation of a molecule.
b. A molecule can have different SMILES representations.
c. A molecular graph can be transformed into a SMILES string and the other way around.

Question 6

In the Q-learning update rule, Q(s,a) ← (1-α)Q(s,a) + α(r + γ maxₐ’ Q(s’,a’)), what is the role of the discount factor γ (gamma)?
a. It determines the learning rate of the update.
b. It determines the importance of future rewards. A value close to 0 makes the agent ‘myopic’ (short-sighted), while a value close to 1 makes it value long-term rewards highly.
c. It controls the trade-off between exploration and exploitation.
d. It ensures that the Q-values do not grow infinitely large.

Answer 6

b. It determines the importance of future rewards. A value close to 0 makes the agent ‘myopic’ (short-sighted), while a value close to 1 makes it value long-term rewards highly.

Question 7

In a standard feed-forward neural network, how is the output of a single neuron calculated?
a. By summing the outputs of all neurons in the previous layer.
b. By calculating a weighted sum of the outputs from the previous layer, adding a bias, and then applying a non-linear activation function.
c. By simply applying a non-linear activation function to the input vector.
d. By performing a convolution operation on the inputs.

Answer 7

b. By calculating a weighted sum of the outputs from the previous layer, adding a bias, and then applying a non-linear activation function.

Question 8

What is the purpose of a pooling layer in a CNN?
a. To increase the spatial dimensions of the feature maps.
b. To progressively reduce the spatial size of the representation.
c. To reduce the number of parameters and computational load in the network.
d. To make the representation more robust to small translations in the input.

Answer 8

b. To progressively reduce the spatial size of the representation.
c. To reduce the number of parameters and computational load in the network.
d. To make the representation more robust to small translations in the input.

Question 9

**Which of the following statements are true about the vanishing gradient problem?
a. Repeated multiplication of gradients smaller than 1 leads to a vanishing gradient.
b. The choice of the activation functions plays a crucial role in the vanishing gradient problem.
c. A vanishing gradient can be mitigated by decreasing the learning rate.
d. It is more severe in the final layers of a deep network (closer to the output).
e. Vanishing gradients make the training of a neural network extremely difficult.

Answer 9

a. Repeated multiplication of gradients smaller than 1 leads to a vanishing gradient.
b. The choice of the activation functions plays a crucial role in the vanishing gradient problem.
e. Vanishing gradients make the training of a neural network extremely difficult.

Question 10

**Which of the following statements is/are true about (deep) Q-learning?
a. Q-learning is one possible implementation of reinforcement learning.
b. Q-learning becomes computationally infeasible for larger MDPs (Markov decision processes) due to the large state-action space.
c. Deep Q-learning is about approximating the Q-value function using a neural network.
d. It is used to learn an optimal policy by estimating Q-values.

Answer 10

a. Q-learning is one possible implementation of reinforcement learning.
b. Q-learning becomes computationally infeasible for larger MDPs (Markov decision processes) due to the large state-action space.
c. Deep Q-learning is about approximating the Q-value function using a neural network.
d. It is used to learn an optimal policy by estimating Q-values.

Question 11

**With respect to the vanishing gradient problem, which of the following statements are true regarding deep neural networks?
a. The deeper the network, the more multiplications (chain rule) we have to perform in the backward pass.
b. The vanishing gradient problem can get more severe when increasing the network depth.
c. The vanishing gradient problem will typically occur towards the input layer.
d. The problem is independent of the network’s depth and only depends on the activation function.

Answer 11

a. The deeper the network, the more multiplications (chain rule) we have to perform in the backward pass.
b. The vanishing gradient problem can get more severe when increasing the network depth.
c. The vanishing gradient problem will typically occur towards the input layer.

Question 12

What is a ‘hyperparameter’ in machine learning?
a. A parameter of the model that is learned during the training process, such as a weight in a neural network.
b. A configuration that is set before the training process begins, such as the learning rate, the number of hidden layers, or the value of k in k-Means.
c. The output prediction of a model.
d. The loss function used to train a model.

Answer 12

b. A configuration that is set before the training process begins, such as the learning rate, the number of hidden layers, or the value of k in k-Means.

Question 13

Which of the following are valid representations for small molecules in cheminformatics?
a. SMILES strings.
b. Molecular graphs, where atoms are nodes and bonds are edges.
c. Molecular fingerprints (e.g., Morgan fingerprints), which are binary vectors.
d. A 3D coordinate list for each atom.

Answer 13

a. SMILES strings.
b. Molecular graphs, where atoms are nodes and bonds are edges.
c. Molecular fingerprints (e.g., Morgan fingerprints), which are binary vectors.
d. A 3D coordinate list for each atom.

Question 14

Which of these statements about the chain rule’s role in backpropagation is correct?
a. The chain rule is used to compute the gradient of a composite function.
b. In deep networks, the gradient of the loss with respect to an early layer’s weights is calculated by multiplying the derivatives of all subsequent layers.
c. This multiplicative nature is what can lead to the vanishing or exploding gradient problems.
d. The chain rule simplifies the computation by breaking down the gradient calculation into a product of local derivatives.

Answer 14

a. The chain rule is used to compute the gradient of a composite function.
b. In deep networks, the gradient of the loss with respect to an early layer’s weights is calculated by multiplying the derivatives of all subsequent layers.
c. This multiplicative nature is what can lead to the vanishing or exploding gradient problems.
d. The chain rule simplifies the computation by breaking down the gradient calculation into a product of local derivatives.

Question 15

The convolution of a 64x64 grayscale image (1 channel) with 16 kernels of size 5x5 (with no padding and a stride of 1) produces…
a. … an output with 1 feature map.
b. … an output with 16 feature maps.
c. … an output volume of size 60x60x16.
d. … an output volume of size 64x64x16.

Answer 15

b. … an output with 16 feature maps.
c. … an output volume of size 60x60x16.

Question 16

Which of the following is true about padding in Convolutional Neural Networks?
a. It is the process of adding extra pixels (usually zeros) around the border of an input image.
b. It can be used to control the spatial size of the output feature maps.
c. ‘Valid’ padding means no padding is applied, which typically causes the output feature map to be smaller than the input.
d. ‘Same’ padding aims to keep the output feature map the same size as the input feature map.

Answer 16

a. It is the process of adding extra pixels (usually zeros) around the border of an input image.
b. It can be used to control the spatial size of the output feature maps.
c. ‘Valid’ padding means no padding is applied, which typically causes the output feature map to be smaller than the input.
d. ‘Same’ padding aims to keep the output feature map the same size as the input feature map.

Question 17

**Which of the following statements is/are true about QSAR (Quantitative Structure-Activity Relationship)?
a. The bio-activity of a molecule is determined by its molecular structure.
b. The hypothesis is that similar molecular structures have similar activities.
c. The hypothesis is that molecules with similar activities must have similar molecular structures.
d. It is primarily used for predicting the cost of drug development.

Answer 17

a. The bio-activity of a molecule is determined by its molecular structure.
b. The hypothesis is that similar molecular structures have similar activities.

Question 18

**The original CEC (constant error carousel) of an LSTM …
a. … is responsible for countering the vanishing gradient problem.
b. … is responsible for going from the old cell state to the new cell state.
c. … uses non-linear functions to modify the cell state.
d. … is the primary cause of the vanishing gradient problem in RNNs.

Answer 18

a. … is responsible for countering the vanishing gradient problem.
b. … is responsible for going from the old cell state to the new cell state.

Question 19

In the self-attention formula Attention(Q, K, V) = softmax(QKᵀ/√dₖ)V, what do Q, K, and V represent?
a. Q, K, and V are learnable weight matrices.
b. Q, K, and V (Query, Key, Value) are vectors derived from the input embeddings for each token.
c. The dot product of a query (Q) and a key (K) determines an attention score, indicating how much focus to place on another token.
d. The value vectors (V) are averaged, weighted by the attention scores, to produce the output for each token.

Answer 19

b. Q, K, and V (Query, Key, Value) are vectors derived from the input embeddings for each token.
c. The dot product of a query (Q) and a key (K) determines an attention score, indicating how much focus to place on another token.
d. The value vectors (V) are averaged, weighted by the attention scores, to produce the output for each token.

Question 20

**Which of the following are parts of a Markov decision process?
a. A set of possible states.
b. A set of possible actions.
c. A set of possible rewards.
d. A set of possible state transitions.
e. A policy function.

Answer 20

a. A set of possible states.
b. A set of possible actions.

Question 21

*Which of the following statements are true regarding a loss function?
a. It measures how close the prediction is to the true target.
b. Typically, the lower the loss, the better the prediction.
c. The choice of loss function depends on the task (e.g., Cross-Entropy for classification, MSE for regression).
d. A loss of 0 indicates a perfect prediction on the given sample.

Answer 21

a. It measures how close the prediction is to the true target.
b. Typically, the lower the loss, the better the prediction.
c. The choice of loss function depends on the task (e.g., Cross-Entropy for classification, MSE for regression).
d. A loss of 0 indicates a perfect prediction on the given sample.

Question 22

What is a ‘receptive field’ in a Convolutional Neural Network?
a. The entire input image that a neuron in the output layer can ‘see’.
b. The specific region of the input volume that a particular neuron in a convolutional layer is connected to.
c. As we go deeper into the network, the effective receptive field of the neurons increases.
d. It is a learnable parameter of the convolutional layer.

Answer 22

b. The specific region of the input volume that a particular neuron in a convolutional layer is connected to.
c. As we go deeper into the network, the effective receptive field of the neurons increases.

Question 23

**Back-propagation through time (BPTT) …
a. … is commonly used to train an RNN.
b. … generates a computational graph that is (potentially) very deep, depending on sequence length.
c. … is the process of calculating the gradient of the loss with respect to an RNN’s weights.
d. … mitigates the vanishing gradient problem by design.

Answer 23

a. … is commonly used to train an RNN.
b. … generates a computational graph that is (potentially) very deep, depending on sequence length.
c. … is the process of calculating the gradient of the loss with respect to an RNN’s weights.

Question 24

What is the primary advantage of the Transformer architecture over RNNs for language tasks?
a. Transformers process the entire input sequence at once, allowing for massive parallelization during training.
b. The self-attention mechanism allows the model to directly weigh the importance of all other words in the sequence when processing a given word.
c. Transformers are inherently better at handling short sequences than RNNs.
d. Transformers have fewer parameters than LSTMs, making them faster to train.

Answer 24

a. Transformers process the entire input sequence at once, allowing for massive parallelization during training.
b. The self-attention mechanism allows the model to directly weigh the importance of all other words in the sequence when processing a given word.

Question 25

What does it mean for a dataset to be 'i.i.d.' (independently and identically distributed)? a. Each data sample is drawn from a different probability distribution. b. Each data sample has the same probability distribution as the others, and all samples are mutually independent. c. This assumption is crucial for the test set to be a good estimator of the generalization error. d. Sequential data, like time series, generally violates the 'i.i.d.' assumption.

Answer 25

b. Each data sample has the same probability distribution as the others, and all samples are mutually independent. c. This assumption is crucial for the test set to be a good estimator of the generalization error. d. Sequential data, like time series, generally violates the 'i.i.d.' assumption.

Question 26

What is the primary function of the decoder in an RNN-based language model? a. To convert an input word into a dense vector embedding. b. To process the sequence of embeddings and maintain a hidden state. c. To take the final hidden state from the RNN and map it to a probability distribution over the entire vocabulary. d. To select the most likely next word from the probability distribution.

Answer 26

c. To take the final hidden state from the RNN and map it to a probability distribution over the entire vocabulary.

Question 27

**Which of the following statements is/are true about language models? a. Language modeling is the task of predicting a word given a context. b. During training of an RNN language model, at each timestep, the decoder provides the probability distribution of the next word over the entire vocabulary. c. Perplexity is a common metric to evaluate language models, where lower is better. d. They can only be built using Recurrent Neural Networks.

Answer 27

a. Language modeling is the task of predicting a word given a context. b. During training of an RNN language model, at each timestep, the decoder provides the probability distribution of the next word over the entire vocabulary. c. Perplexity is a common metric to evaluate language models, where lower is better.

Question 28

Why is a non-linear activation function crucial in a multi-layer neural network? a. Without non-linearity, stacking multiple layers is equivalent to a single linear layer. b. They make the backpropagation process computationally cheaper. c. They allow the network to learn complex, non-linear relationships in the data. d. They ensure that the output of the network is always between 0 and 1.

Answer 28

a. Without non-linearity, stacking multiple layers is equivalent to a single linear layer. c. They allow the network to learn complex, non-linear relationships in the data.

Question 29

Which statement best describes the difference between a model's parameters and its hyperparameters? a. Parameters are learned from the data (e.g., weights), while hyperparameters are set by the user before training (e.g., learning rate). b. Hyperparameters are learned from the data, while parameters are set by the user. c. Both are learned from the data, but hyperparameters are in the final layer. d. There is no fundamental difference; the terms are interchangeable.

Answer 29

a. Parameters are learned from the data (e.g., weights), while hyperparameters are set by the user before training (e.g., learning rate).

Question 30

What is the main purpose of dimensionality reduction techniques like PCA and t-SNE? a. To increase the number of features in a dataset. b. To visualize high-dimensional data in a lower-dimensional space (e.g., 2D or 3D). c. To reduce the computational complexity of machine learning algorithms. d. To preserve as much of the data's variance or structure as possible while reducing dimensions. e. To always improve the performance of a classifier.

Answer 30

b. To visualize high-dimensional data in a lower-dimensional space (e.g., 2D or 3D). c. To reduce the computational complexity of machine learning algorithms. d. To preserve as much of the data's variance or structure as possible while reducing dimensions.

Question 31

**Which of the following statements are true about data augmentation? a. It can help prevent overfitting. b. It might increase the robustness of a model. c. Some data augmentation techniques (e.g., adding noise) can be applied to various data types. d. It always improves model performance.

Answer 31

a. It can help prevent overfitting. b. It might increase the robustness of a model. c. Some data augmentation techniques (e.g., adding noise) can be applied to various data types.

Question 32

In the context of the basic data analysis workflow, why is a held-out test set important? a. It is used to tune the model's hyperparameters. b. It provides an unbiased estimate of the model's performance on unseen data (generalization error). c. The model should only be evaluated on the test set once, after all training and hyperparameter tuning is complete. d. Repeatedly evaluating on the test set and choosing the model that performs best on it can lead to overfitting to the test set.

Answer 32

b. It provides an unbiased estimate of the model's performance on unseen data (generalization error). c. The model should only be evaluated on the test set once, after all training and hyperparameter tuning is complete. d. Repeatedly evaluating on the test set and choosing the model that performs best on it can lead to overfitting to the test set.

Question 33

What does 'overfitting' mean in the context of supervised learning? a. The model performs very well on the training data but poorly on unseen test data. b. The model has learned the noise and specific artifacts of the training data, rather than the underlying general pattern. c. A highly complex model is more prone to overfitting than a simpler model. d. Overfitting can be detected by comparing the performance on the training set with the performance on a validation or test set.

Answer 33

a. The model performs very well on the training data but poorly on unseen test data. b. The model has learned the noise and specific artifacts of the training data, rather than the underlying general pattern. c. A highly complex model is more prone to overfitting than a simpler model. d. Overfitting can be detected by comparing the performance on the training set with the performance on a validation or test set.

Question 34

Why might a learning rate schedule be beneficial during training? a. Using a large learning rate initially can help escape sharp, poor local minima and make faster progress. b. Decreasing the learning rate over time allows the model to settle into a good minimum more carefully. c. It helps to prevent the training from getting stuck with a learning rate that is too high or too low. d. A fixed learning rate is always optimal for all stages of training.

Answer 34

a. Using a large learning rate initially can help escape sharp, poor local minima and make faster progress. b. Decreasing the learning rate over time allows the model to settle into a good minimum more carefully. c. It helps to prevent the training from getting stuck with a learning rate that is too high or too low.

Question 35

In the context of the basic data analysis workflow, what is 'preprocessing'? a. The final step where the model's answer is interpreted. b. An initial step that involves cleaning, normalizing, and transforming the raw data to make it suitable for a machine learning model. c. The process of choosing the model class (e.g., SVM, Neural Network). d. The process of training the model on the data.

Answer 35

b. An initial step that involves cleaning, normalizing, and transforming the raw data to make it suitable for a machine learning model.

Question 36

**You are given the derivative of the logistic/sigmoid function σ'(x) and the following computation: y = σ'(10) * σ'(-10). Which of the following statements is/are correct? a. y is very close to 0. b. y is in the range (0, 1/16]. c. y is a positive value. d. y is exactly 0.

Answer 36

a. y is very close to 0. b. y is in the range (0, 1/16]. c. y is a positive value.

Question 37

In Reinforcement Learning, what is the difference between model-based and model-free approaches? a. Model-free methods (like Q-learning) learn a policy or value function directly from experience without explicitly learning the environment's dynamics (transition probabilities and rewards). b. Model-based methods first try to learn a model of the environment and then use this model for planning (e.g., by simulating future trajectories). c. Model-based methods are generally more sample-efficient but can suffer if the learned model of the environment is inaccurate. d. There is no difference; all RL algorithms learn a model of the environment.

Answer 37

a. Model-free methods (like Q-learning) learn a policy or value function directly from experience without explicitly learning the environment's dynamics (transition probabilities and rewards). b. Model-based methods first try to learn a model of the environment and then use this model for planning (e.g., by simulating future trajectories). c. Model-based methods are generally more sample-efficient but can suffer if the learned model of the environment is inaccurate.

Question 38

Which of the following statements about word embeddings are true? a. They represent words as dense, low-dimensional vectors. b. Words with similar meanings are expected to have similar vectors (i.e., be close in the vector space). c. Cosine similarity is a common metric to measure the similarity between two word vectors. d. One-hot encoding is a form of dense word embedding.

Answer 38

a. They represent words as dense, low-dimensional vectors. b. Words with similar meanings are expected to have similar vectors (i.e., be close in the vector space). c. Cosine similarity is a common metric to measure the similarity between two word vectors.

Question 39

In a Transformer model, what is the purpose of positional encodings? a. To introduce non-linearity into the model. b. Since the self-attention mechanism does not inherently process sequential order, positional encodings are added to the input embeddings to give the model information about the position of tokens in the sequence. c. They are learnable parameters that are updated during training. d. They are fixed vectors calculated using sine and cosine functions of different frequencies.

Answer 39

b. Since the self-attention mechanism does not inherently process sequential order, positional encodings are added to the input embeddings to give the model information about the position of tokens in the sequence. d. They are fixed vectors calculated using sine and cosine functions of different frequencies.

Question 40

Which of the following are potential solutions to the vanishing gradient problem in deep networks? a. Using ReLU or its variants (like Leaky ReLU) as activation functions. b. Using residual connections (as in ResNets). c. Using Batch Normalization. d. Using a very small, fixed learning rate. e. Using LSTM or GRU cells in recurrent networks.

Answer 40

a. Using ReLU or its variants (like Leaky ReLU) as activation functions. b. Using residual connections (as in ResNets). c. Using Batch Normalization. e. Using LSTM or GRU cells in recurrent networks.

Question 41

**The 3D structure of a molecule ... a. ... determines its functionality to a high degree. b. ... is straightforward to predict if the sequence of atoms is known. c. ... is expensive to determine experimentally. d. ... can be different for the same molecular graph. e. ... is unambiguously determined by a molecular graph.

Answer 41

a. ... determines its functionality to a high degree. c. ... is expensive to determine experimentally. d. ... can be different for the same molecular graph.

Question 42

Which of the following are unsupervised learning tasks? a. k-Means Clustering b. Linear Regression c. Principal Component Analysis (PCA) d. Classifying images of cats and dogs. e. Density Based Clustering (DBSCAN)

Answer 42

a. k-Means Clustering c. Principal Component Analysis (PCA) e. Density Based Clustering (DBSCAN)

Question 43

The protein folding problem, famously tackled by AlphaFold, is about: a. Predicting the amino acid sequence of a protein from its 3D structure. b. Predicting the 3D structure of a protein from its amino acid sequence. c. Synthesizing new proteins in a lab. d. Finding which drugs will bind to a folded protein.

Answer 43

b. Predicting the 3D structure of a protein from its amino acid sequence.

Question 44

What is the core principle of weight sharing in Convolutional Neural Networks (CNNs)? a. Each neuron in a feature map shares its weights with neurons in different feature maps. b. The same set of weights (kernel/filter) is applied across different locations of the input image. c. It drastically reduces the number of learnable parameters compared to a fully-connected network. d. It helps the network to detect the same feature (e.g., an edge) regardless of its position in the image, promoting translation invariance.

Answer 44

b. The same set of weights (kernel/filter) is applied across different locations of the input image. c. It drastically reduces the number of learnable parameters compared to a fully-connected network. d. It helps the network to detect the same feature (e.g., an edge) regardless of its position in the image, promoting translation invariance.

Question 45

What is the primary motivation for using Batch Normalization? a. To reduce the effect of internal covariate shift, which is the change in the distribution of network activations due to changes in network parameters during training. b. To act as a strong regularizer, often reducing the need for Dropout. c. To allow for the use of much higher learning rates and accelerate training. d. To ensure all weights in the network remain positive.

Answer 45

a. To reduce the effect of internal covariate shift, which is the change in the distribution of network activations due to changes in network parameters during training. b. To act as a strong regularizer, often reducing the need for Dropout. c. To allow for the use of much higher learning rates and accelerate training.

Question 46

**Which of the following statements are true about the weights of an RNN? a. For each timestep, the same (shared) weight matrix is used. b. The size of the weight matrix is independent of the sequence length. c. For each timestep, a dedicated, new weight matrix is learned. d. For sequences of different length, multiple (shared) weight matrices must be learned (one for each unique sequence length).

Answer 46

a. For each timestep, the same (shared) weight matrix is used. b. The size of the weight matrix is independent of the sequence length.

Question 47

What distinguishes Density-Based Spatial Clustering of Applications with Noise (DBSCAN) from k-Means? a. DBSCAN does not require the user to specify the number of clusters beforehand. b. DBSCAN can find arbitrarily shaped clusters. c. DBSCAN can identify points as outliers (noise). d. DBSCAN is a parametric clustering algorithm.

Answer 47

a. DBSCAN does not require the user to specify the number of clusters beforehand. b. DBSCAN can find arbitrarily shaped clusters. c. DBSCAN can identify points as outliers (noise).

Question 48

*In regression, ... a. ... the target value is numeric. b. ... we are dealing with a supervised learning scenario. c. ... Mean Squared Error (MSE) is a common loss function. d. ... the output of the model is a continuous value.

Answer 48

a. ... the target value is numeric. b. ... we are dealing with a supervised learning scenario. c. ... Mean Squared Error (MSE) is a common loss function. d. ... the output of the model is a continuous value.

Question 49

What is the 'exploration vs. exploitation' trade-off in reinforcement learning? a. The dilemma of choosing between actions that have been effective in the past (exploitation) and trying new actions to discover potentially better rewards (exploration). b. An epsilon-greedy strategy is a common way to manage this trade-off. c. Pure exploitation can lead to getting stuck in a suboptimal policy. d. Pure exploration means acting randomly and may prevent the agent from learning an effective policy.

Answer 49

a. The dilemma of choosing between actions that have been effective in the past (exploitation) and trying new actions to discover potentially better rewards (exploration). b. An epsilon-greedy strategy is a common way to manage this trade-off. c. Pure exploitation can lead to getting stuck in a suboptimal policy. d. Pure exploration means acting randomly and may prevent the agent from learning an effective policy.

Question 50

*Which of the following statements are true about text generation with language models? a. Given a starting word, always choosing the next word with the highest predicted probability will always yield the same generated text. b. Sampling from the predicted output probability distribution can introduce diversity in the generated text. c. A "temperature" parameter can be used to control the randomness of the sampling process. d. The language model needs to be retrained for every new starting word.

Answer 50

a. Given a starting word, always choosing the next word with the highest predicted probability will always yield the same generated text. b. Sampling from the predicted output probability distribution can introduce diversity in the generated text. c. A "temperature" parameter can be used to control the randomness of the sampling process.

Question 51

*Which of the following statements are true about many-to-many RNN types? a. Given an input of length Tx, the output/prediction length can be equal, i.e., Tx = Ty. b. Given an input of length Tx, the output/prediction length can be unequal, i.e., Tx != Ty. c. Machine translation is an example where Tx and Ty can be different. d. Named Entity Recognition is an example where Tx and Ty are typically equal.

Answer 51

a. Given an input of length Tx, the output/prediction length can be equal, i.e., Tx = Ty. b. Given an input of length Tx, the output/prediction length can be unequal, i.e., Tx != Ty. c. Machine translation is an example where Tx and Ty can be different. d. Named Entity Recognition is an example where Tx and Ty are typically equal.

Question 52

What problem does subword tokenization (like BPE or SentencePiece) solve? a. It helps handle rare or out-of-vocabulary (OOV) words by breaking them down into smaller, known units. b. It allows the model to have a fixed-size vocabulary while still being able to represent an open-ended set of words. c. It can capture morphological similarities between words (e.g., 'running' and 'ran' might share the subword 'run'). d. It always results in longer sequences than word-level tokenization.

Answer 52

a. It helps handle rare or out-of-vocabulary (OOV) words by breaking them down into smaller, known units. b. It allows the model to have a fixed-size vocabulary while still being able to represent an open-ended set of words. c. It can capture morphological similarities between words (e.g., 'running' and 'ran' might share the subword 'run').

Question 53

What are molecular fingerprints? a. A sequence representation of a molecule like SMILES. b. A method to represent molecular structures as binary vectors. c. Each bit in the vector typically corresponds to the presence or absence of a specific substructure or chemical feature. d. They are a form of molecular descriptor used as input for machine learning models.

Answer 53

b. A method to represent molecular structures as binary vectors. c. Each bit in the vector typically corresponds to the presence or absence of a specific substructure or chemical feature. d. They are a form of molecular descriptor used as input for machine learning models.

Question 54

**In classification, ... a. ... the target value is a class label. b. ... we are dealing with a supervised machine learning scenario. c. ... the target can be a number (e.g., 0, 1, 2 for different classes). d. ... there can only be two classes.

Answer 54

a. ... the target value is a class label. b. ... we are dealing with a supervised machine learning scenario. c. ... the target can be a number (e.g., 0, 1, 2 for different classes).

Question 55

**What is the general data flow through an RNN language model? a. input -> encoder -> RNN -> decoder b. input -> decoder -> RNN -> encoder c. input -> RNN -> encoder -> decoder d. input -> encoder -> decoder -> RNN

Answer 55

a. input -> encoder -> RNN -> decoder

Question 56

**Which of the following statements is/are true about the ReLU activation function? a. Its derivative is 0 for negative inputs. b. Its derivative is 1 for positive inputs. c. It can help mitigate the vanishing gradient problem. d. For positive inputs, ReLU equates to the identity function. e. It is computationally more expensive than the sigmoid function.

Answer 56

a. Its derivative is 0 for negative inputs. b. Its derivative is 1 for positive inputs. c. It can help mitigate the vanishing gradient problem. d. For positive inputs, ReLU equates to the identity function.

Question 57

**A non-convex function... a. ... is common in deep learning loss landscapes. b. ... often requires iterative methods like gradient descent to find a minimum. c. ... might have several local minima. d. ... typically does not have a closed-form solution for finding the minimum.

Answer 57

a. ... is common in deep learning loss landscapes. b. ... often requires iterative methods like gradient descent to find a minimum. c. ... might have several local minima. d. ... typically does not have a closed-form solution for finding the minimum.

Question 58

What are 'residual connections' as used in networks like ResNet? a. They are connections that add the input of a layer (or block of layers) to its output. b. They create a 'shortcut' for the gradient to flow through during backpropagation. c. They help in training very deep neural networks by mitigating the vanishing gradient problem. d. They are a form of data augmentation.

Answer 58

a. They are connections that add the input of a layer (or block of layers) to its output. b. They create a 'shortcut' for the gradient to flow through during backpropagation. c. They help in training very deep neural networks by mitigating the vanishing gradient problem.

Question 59

**Which of the following statements are true about virtual screening (VS)? a. In VS, a trained neural network predicts whether a molecule from a database is likely to bind/react with the target of interest. b. VS helps to narrow down the number of candidates for expensive experimental testing (assays). c. It is a computational technique used in the early stages of drug discovery. d. VS relies on the QSAR principle.

Answer 59

a. In VS, a trained neural network predicts whether a molecule from a database is likely to bind/react with the target of interest. b. VS helps to narrow down the number of candidates for expensive experimental testing (assays). c. It is a computational technique used in the early stages of drug discovery. d. VS relies on the QSAR principle.

Question 60

*Which of the following statements are true about a standard, single-layer RNN? a. The input sequence must be fed timestep by timestep to the RNN. b. The hidden state is updated at each timestep. c. It can struggle with long-term dependencies due to vanishing gradients. d. Unrolling means representing the recurrent computation over time as a deep feed-forward network.

Answer 60

a. The input sequence must be fed timestep by timestep to the RNN. b. The hidden state is updated at each timestep. c. It can struggle with long-term dependencies due to vanishing gradients. d. Unrolling means representing the recurrent computation over time as a deep feed-forward network.

Question 61

In an LSTM, what is the role of the forget gate? a. It decides what new information to store in the cell state. b. It decides what information from the previous cell state should be discarded. c. It decides what part of the cell state should be used to compute the hidden state output. d. It is responsible for resetting the entire cell state to zero at every timestep.

Answer 61

b. It decides what information from the previous cell state should be discarded.

Question 62

*Which of the following activation functions are prone to the vanishing gradient problem? a. Sigmoid b. Tanh c. ReLU d. Leaky ReLU

Answer 62

a. Sigmoid b. Tanh

Question 63

*Which of the following activation functions can be used to mitigate the vanishing gradient problem? a. ReLU b. Leaky ReLU c. Sigmoid d. SELU

Answer 63

a. ReLU b. Leaky ReLU d. SELU

Question 64

In the context of the RGB color model, what does an 8-bit color depth signify? a. Each of the three channels (Red, Green, Blue) can take on one of 8 possible values. b. The total number of possible colors is 8. c. Each pixel is represented by 8 bits in total. d. Each of the three channels (Red, Green, Blue) is represented by 8 bits, allowing for 256 different intensity values per channel.

Answer 64

d. Each of the three channels (Red, Green, Blue) is represented by 8 bits, allowing for 256 different intensity values per channel.

Question 65

When training a neural network, what does a single 'epoch' refer to? a. A single pass of the gradient descent algorithm over one batch of data. b. One complete forward and backward pass for a single training example. c. One complete pass through the entire training dataset. d. The process of evaluating the model on the test set.

Answer 65

c. One complete pass through the entire training dataset.

Question 66

*Which of the following statements are true when comparing an LSTM to a standard recurrent neural network? a. An LSTM is better able to learn long-term dependencies. b. An LSTM is less prone to the vanishing gradient problem. c. An LSTM has more model parameters. d. An LSTM is computationally more expensive.

Answer 66

a. An LSTM is better able to learn long-term dependencies. b. An LSTM is less prone to the vanishing gradient problem. c. An LSTM has more model parameters. d. An LSTM is computationally more expensive.

Question 67

**Which of the following statements is/are true about gating/gates in an LSTM? a. Gating is used to control what to read/write/forget in memory. b. Typically, the sigmoid function is used as the activation for gates to produce values between 0 and 1. c. Gates depend on the current input and the hidden state from the previous timestep. d. The forget gate was part of the original LSTM design by Hochreiter & Schmidhuber.

Answer 67

a. Gating is used to control what to read/write/forget in memory. b. Typically, the sigmoid function is used as the activation for gates to produce values between 0 and 1. c. Gates depend on the current input and the hidden state from the previous timestep.

Question 68

Which of the following statements correctly describe k-Means clustering? a. It is a supervised learning algorithm. b. The user must specify the number of clusters (k) beforehand. c. The algorithm iteratively assigns data points to the nearest cluster centroid and then recalculates the centroid. d. It is guaranteed to find the globally optimal clustering. e. It is a non-parametric clustering algorithm.

Answer 68

b. The user must specify the number of clusters (k) beforehand. c. The algorithm iteratively assigns data points to the nearest cluster centroid and then recalculates the centroid. e. It is a non-parametric clustering algorithm.

Question 69

**Which of the following statements are true about sequence data input and standard, feed-forward neural networks? a. The relation of elements in the sequence is lost. b. Sequences with variable lengths must be preprocessed (e.g., padded or truncated) to have a fixed size. c. A sliding window approach can be used to capture some local sequential information. d. A feed-forward network treats the input as an unordered set of features.

Answer 69

a. The relation of elements in the sequence is lost. b. Sequences with variable lengths must be preprocessed (e.g., padded or truncated) to have a fixed size. c. A sliding window approach can be used to capture some local sequential information.

Question 70

How does Dropout work as a regularization technique? a. During training, it permanently removes a fixed percentage of neurons from the network. b. During training, it randomly sets the output of a fraction of neurons to zero for each training example. c. It forces the network to learn more robust features that are not dependent on the presence of specific other neurons. d. At test time, all neurons are used, but their weights are scaled down to compensate for the effect of dropout during training.

Answer 70

b. During training, it randomly sets the output of a fraction of neurons to zero for each training example. c. It forces the network to learn more robust features that are not dependent on the presence of specific other neurons. d. At test time, all neurons are used, but their weights are scaled down to compensate for the effect of dropout during training.

Question 71

**Which of the following statements are true about recurrent neural networks? a. They are suitable for processing sequences of variable length. b. They are suitable for processing sequences of fixed length. c. They incorporate information from previous timesteps/positions of a sequence via the hidden state. d. The input sequence must be fed timestep by timestep to the RNN.

Answer 71

a. They are suitable for processing sequences of variable length. b. They are suitable for processing sequences of fixed length. c. They incorporate information from previous timesteps/positions of a sequence via the hidden state. d. The input sequence must be fed timestep by timestep to the RNN.

Question 72

Which of these statements accurately describes the difference between L1 and L2 regularization? a. L1 regularization adds the sum of the absolute values of the weights to the loss function. b. L2 regularization adds the sum of the squared values of the weights to the loss function. c. L1 regularization can lead to sparse models, where some weights become exactly zero. d. L2 regularization is also known as 'weight decay'.

Answer 72

a. L1 regularization adds the sum of the absolute values of the weights to the loss function. b. L2 regularization adds the sum of the squared values of the weights to the loss function. c. L1 regularization can lead to sparse models, where some weights become exactly zero. d. L2 regularization is also known as 'weight decay'.

Question 73

What is Transfer Learning in the context of deep learning? a. Training a model from scratch on a new, small dataset. b. Taking a model pre-trained on a large dataset (e.g., ImageNet) and adapting it for a new, related task. c. Freezing the weights of the early convolutional layers and only training the final classification layers on the new task. d. It is a technique that generally requires less data and less computation time for the new task compared to training from scratch.

Answer 73

b. Taking a model pre-trained on a large dataset (e.g., ImageNet) and adapting it for a new, related task. c. Freezing the weights of the early convolutional layers and only training the final classification layers on the new task. d. It is a technique that generally requires less data and less computation time for the new task compared to training from scratch.

Question 74

**Which of the following statements are true about the concept of an optimal policy in Reinforcement Learning? a. It maximizes the sum of rewards over an episode. b. There can be multiple optimal policies for a single MDP. c. It can be found using Q-learning. d. An optimal policy is one that always chooses the action with the highest immediate reward.

Answer 74

a. It maximizes the sum of rewards over an episode. b. There can be multiple optimal policies for a single MDP. c. It can be found using Q-learning.