Question 1

When building a deep neural network from scratch, which process is primarily responsible for updating the model's weights by calculating the gradient of the loss function with respect to each weight using the chain rule?

Accepted Answer

Backpropagation

Answer

Forward propagation

Answer

Data augmentation

Answer

Feature engineering

Question 2

In the context of generative modelling, which of the following describes the 'density estimation' challenge when dealing with unstructured data like images or audio?

Accepted Answer

Learning the underlying probability distribution that governs the training data to generate new, realistic samples.

Answer

Ensuring that the model only memorises the training set to prevent it from creating unknown variations.

Answer

Converting all unstructured data into a tabular format with predefined headers before training the network.

Answer

Manually labelling every pixel in an image to ensure the model understands the physical objects present.

Question 3

When improving a Deep Neural Network that exhibits high variance (overfitting), which technique is most appropriate to simplify the model's complexity and improve its generalisation on unseen data?

Accepted Answer

Regularisation

Answer

Increasing the number of layers

Answer

One-hot encoding

Answer

Stochastic Gradient Descent

Question 4

In the context of probabilistic generative models, what is the primary distinction between the density estimation approach and the sample generation approach?

Accepted Answer

Density estimation explicitly determines the probability density function

p(x)

, whereas sample generation focuses on producing new data points without necessarily uncovering the explicit density.

Answer

Density estimation only works with structured data, while sample generation only works with unstructured data.

Answer

Density estimation requires a supervised label for every input, whereas sample generation is purely an unsupervised task.

Answer

Density estimation is used to improve the model's speed, while sample generation is used to improve the model's memory efficiency.

Question 5

When creating a deep neural network from scratch, which activation function is commonly used in the hidden layers to mitigate the vanishing gradient problem, expressed as

p(x)

?

Accepted Answer

Rectified Linear Unit (ReLU)

Answer

Hyperbolic Tangent (tanh)

Answer

Sigmoid

Answer

Softmax

Question 6

In the context of deep learning and data types, what is a primary structural difference between unstructured data and structured data that a Deep Neural Network must address?

Accepted Answer

Unstructured data lacks a predefined data model or schema, requiring the network to automatically extract complex features.

Answer

Unstructured data can only be processed by shallow networks, whereas structured data requires deep architectures.

Answer

Structured data always contains more information per byte than unstructured data.

Answer

Structured data is stored as binary large objects (BLOBs) while unstructured data is stored in relational tables.

Question 7

Which of the following describes the 'curse of dimensionality', a significant challenge in generative modelling for deep learning?

Accepted Answer

As the number of features or dimensions increases, the amount of data needed to effectively sample the space increases exponentially.

Answer

The model becomes too fast to converge, causing it to skip the global minimum of the loss function.

Answer

Data points in high-dimensional space become inherently closer to each other, making clustering too easy.

Answer

The neural network layers become physically too large to fit into the memory of a standard GPU.

Question 8

When improving a model by selecting an appropriate loss function for a multi-class classification problem within a deep neural network, which function is typically combined with a Softmax output layer?

Accepted Answer

Categorical Cross-Entropy

Answer

Binary Cross-Entropy

Answer

Huber Loss

Answer

Mean Squared Error

Question 9

Which component of a Deep Neural Network, often adjusted to improve the model, is responsible for providing the initial non-linearity that allows the network to approximate complex, non-linear functions?

Accepted Answer

Activation Function

Answer

Learning Rate

Answer

Loss Function

Answer

Weight Initialisation

Question 10

In the process of improving a Deep Neural Network from scratch, what is the primary purpose of using 'Mini-batch Gradient Descent' instead of 'Stochastic Gradient Descent' (using one sample at a time)?

Accepted Answer

To reduce the variance of the parameter updates and utilise highly optimised matrix operations on modern hardware.

Answer

To ensure that the model never reaches a local minimum by increasing the randomness of every single update step.

Answer

To completely eliminate the need for a learning rate hyperparameter during the training phase.

Answer

To transform the unstructured input data into a structured format automatically before the first hidden layer.

Question 11

When constructing a Deep Neural Network from scratch to address the challenges of generative modelling, which technique is most commonly used to mitigate the 'Vanishing Gradient' problem during the model improvement phase?

Accepted Answer

Implementing Batch Normalisation and using non-saturating activation functions like ReLU

Answer

Increasing the number of layers indefinitely without changing the activation functions

Answer

Switching exclusively to structured data formats to simplify the manifold

Answer

Decreasing the learning rate to

p(x)

to ensure stability

Question 12

In the context of probabilistic generative modelling for unstructured data, what is the primary objective of the model during the training phase?

Accepted Answer

To learn the underlying probability distribution

p(x)

of the input data to generate new, similar samples

Answer

To identify the linear correlation between independent variables in a structured database table

Answer

To map high-dimensional input data to a single categorical label with

p(x)

accuracy

Answer

To reduce the dimensionality of data solely for the purpose of storage efficiency and compression

Question 13

When building a Deep Neural Network from scratch, which method is primarily used to refine the model's parameters by calculating the partial derivative of the loss function

p(x)

with respect to each weight

w

?

Accepted Answer

Backpropagation using the Chain Rule

Answer

Linear Regression Interpolation

Answer

Stochastic Grid Searching

Answer

Forward Propagation of Residuals

Question 14

In the context of generative modelling, how does the complexity of 'Unstructured Data' typically differ from 'Structured Data' when designing a Deep Neural Network from scratch?

Accepted Answer

Unstructured data lacks a pre-defined data model, requiring the network to learn hierarchical spatial or temporal features automatically.

Answer

The density of unstructured data is always lower, meaning the model requires fewer parameters to capture the data distribution.

Answer

Structured data requires the use of Latent Variable Models, whereas unstructured data can only be processed by linear regression.

Answer

Unstructured data is organised in rigid tables, making it easier for simple perceptrons to process without hidden layers.

Question 15

One of the fundamental challenges in Generative Modelling is 'Density Estimation'. Within the framework of Probabilistic Generative Models, what does explicit density estimation entail?

Accepted Answer

Defining and solving for an explicit mathematical function that describes the probability density

p(x)

Answer

Increasing the size of the hidden layers until the loss function reaches exactly zero

Answer

Using a discriminator to distinguish between real and fake samples without a likelihood function

Answer

Manually mapping every unstructured data point to a unique integer in a structured array

Question 16

During the process of improving a Deep Neural Network built from scratch, what is the primary purpose of 'Weight Initialisation' strategies, such as Xavier (Glorot) or He initialisation?

Accepted Answer

To prevent signals from becoming too small or too large as they pass through multiple layers by maintaining the variance of activations

Answer

To ensure that the loss function

p(x)

is always equal to zero at the start of training

Answer

To convert unstructured data into a structured format before the first forward pass

Answer

To eliminate the need for an activation function in the hidden layers

Question 17

When designing a Probabilistic Generative Model, what is meant by the 'manifold hypothesis' regarding the distribution of unstructured data?

Accepted Answer

The assumption that high-dimensional data, such as images, actually lies on a lower-dimensional, non-linear subspace.

Answer

The process of converting a generative model into a discriminative model by removing the latent space.

Answer

The requirement that deep neural networks must use at least

p(x)

layers to model structured data.

Answer

The theory that all data points are uniformly distributed across the entire hyper-volume of the input space.

Question 18

When building a Deep Neural Network from scratch, what is the primary purpose of the 'Activation Function' within each neuron, and why is its non-linearity crucial for generative modelling?

Accepted Answer

To allow the network to approximate complex, non-linear mappings between the latent space and high-dimensional data.

Answer

To simplify the model so that it only computes linear combinations of its inputs.

Answer

To ensure that the weights

p(x)

are always updated to a value of zero during backpropagation.

Answer

To transform structured data into unstructured formats for easier density estimation.

Question 19

In the context of 'Improving the Model' after building a Deep Neural Network from scratch, what is the primary role of 'Regularisation' techniques like L2 weight decay or Dropout?

Accepted Answer

To reduce overfitting by penalising large coefficients or randomly disabling neurons, ensuring the model generalises better to unseen data.

Answer

To accelerate the forward propagation speed of unstructured data through the network layers.

Answer

To convert a probabilistic generative model into a deterministic lookup table for structured data.

Answer

To artificially increase the dimensionality of the hidden layers to match the input data exactly.

Question 20

When addressing the 'Challenges of Generative Modelling', what is meant by the 'Sharpness' problem often observed in models like Variational Autoencoders (VAEs) and Generative Adversarial Networks (GANs)?

Accepted Answer

The difficulty in producing high-fidelity, crisp images instead of blurry reconstructions or those with visual artefacts.

Answer

The process of reducing the number of layers in a Deep Neural Network to increase execution speed.

Answer

The mathematical requirement that all activation functions must have a sharp discontinuous peak.

Answer

The issue where the gradient

p(x)

becomes too large for structured data arrays.

Related quizzes

Related quizzes