17/02/2025, 03:02 awesome-generative-ai-guide/interview_prep/60_gen_ai_questions.

md at main · aishwaryanr/awesome-generative-ai-guide · GitHub

aishwaryanr/awesome-generative-ai-guide Public

Code Issues Pull requests Actions Projects Security Insights

awesome-generative-ai-guide / interview_prep / 60_gen_ai_questions.md

aishwaryanr folder restructure 2ba0411 · last year

1014 lines (551 loc) · 95.1 KB

Preview Code Blame Raw

Common Generative AI Interview

Questions
Owner: Aishwarya Nr
Generative Models
1. What is the difference between generative and discriminative models?
Answer:
Generative models, such as Variational Autoencoders (VAEs) and Generative
Adversarial Networks (GANs), are designed to generate new data samples
by understanding and capturing the underlying data distribution.
Discriminative models, on the other hand, focus on distinguishing between
different classes or categories within the data.

[Link] 1/47
17/02/2025, 03:02 awesome-generative-ai-guide/interview_prep/60_gen_ai_questions.md at main · aishwaryanr/awesome-generative-ai-guide · GitHub

Image Source: [Link]

discriminative-models-d8958b67ad32

2. Describe the architecture of a Generative Adversarial Network and how the

generator and discriminator interact during training.
Answer:
A Generative Adversarial Network comprises a generator and a discriminator. The
generator produces synthetic data, attempting to mimic real data, while the
discriminator evaluates the authenticity of the generated samples. During
training, the generator and discriminator engage in a dynamic interplay, each
striving to outperform the other. The generator aims to create more realistic data,
and the discriminator seeks to improve its ability to differentiate between real and
generated samples.

Image Source: [Link]

identification-model/

3. Explain the concept of a Variational Autoencoder (VAE) and how it

incorporates latent variables into its architecture.
Answer:
A Variational Autoencoder (VAE) is a type of neural network architecture used for
unsupervised learning of latent representations of data. It consists of an encoder
and a decoder network.

[Link] 2/47
17/02/2025, 03:02 awesome-generative-ai-guide/interview_prep/60_gen_ai_questions.md at main · aishwaryanr/awesome-generative-ai-guide · GitHub

Image source: [Link]

autoencoders-vaes-f70510919f73
The encoder takes input data and maps it to a probability distribution in a latent
space. Instead of directly producing a single latent vector, the encoder outputs
parameters of a probability distribution, typically Gaussian, representing the
uncertainty in the latent representation. This stochastic process allows for
sampling from the latent space.
The decoder takes these sampled latent vectors and reconstructs the input data.
During training, the VAE aims to minimize the reconstruction error between the
input data and the decoded output, while also minimizing the discrepancy
between the learned latent distribution and a pre-defined prior distribution, often
a standard Gaussian.
By incorporating latent variables into its architecture, the VAE learns a compact
and continuous representation of the input data in the latent space. This enables
meaningful interpolation and generation of new data samples by sampling from
the learned latent distribution. Additionally, the probabilistic nature of the VAE's
latent space allows for uncertainty estimation in the generated outputs.
read this article: [Link]
autoencoders-vaes-f70510919f73

4. How do conditional generative models differ from unconditional ones? Provide an

example scenario where a conditional approach is beneficial.
Answer:
Conditional generative models differ from unconditional ones by considering
additional information or conditions during the generation process. In
unconditional generative models, such as vanilla GANs or VAEs, the model learns
to generate samples solely based on the underlying data distribution. However, in
conditional generative models, the generation process is conditioned on
additional input variables or labels.
[Link] 3/47
17/02/2025, 03:02 awesome-generative-ai-guide/interview_prep/60_gen_ai_questions.md at main · aishwaryanr/awesome-generative-ai-guide · GitHub

For example, in the context of image generation, an unconditional generative

model might learn to generate various types of images without any specific
constraints. On the other hand, a conditional generative model could be trained
to generate images of specific categories, such as generating images of different
breeds of dogs based on input labels specifying the breed.
A scenario where a conditional approach is beneficial is in tasks where precise
control over the generated outputs is required or when generating samples
belonging to specific categories or conditions. For instance:
In image-to-image translation tasks, where the goal is to convert images
from one domain to another (e.g., converting images from day to night), a
conditional approach allows the model to learn the mapping between input
and output domains based on paired data.
In text-to-image synthesis, given a textual description, a conditional
generative model can generate corresponding images that match the
description, enabling applications like generating images from textual
prompts.
Conditional generative models offer greater flexibility and control over the
generated outputs by incorporating additional information or conditions, making
them well-suited for tasks requiring specific constraints or tailored generation
based on input conditions.

5. What is mode collapse in the context of GANs, and what strategies can be
employed to address it during training?
Answer:
Mode collapse in the context of Generative Adversarial Networks (GANs) refers
to a situation where the generator produces limited diversity in generated
samples, often sticking to a few modes or patterns in the data distribution.
Instead of capturing the full richness of the data distribution, the generator might
only learn to generate samples that belong to a subset of the possible modes,
resulting in repetitive or homogeneous outputs.
Several strategies can be employed to address mode collapse during training:
i. Architectural Modifications: Adjusting the architecture of the GAN can help
mitigate mode collapse. This might involve increasing the capacity of the
generator and discriminator networks, introducing skip connections, or
employing more complex network architectures such as deep convolutional
GANs (DCGANs) or progressive growing GANs (PGGANs).
[Link] 4/47
17/02/2025, 03:02 awesome-generative-ai-guide/interview_prep/60_gen_ai_questions.md at main · aishwaryanr/awesome-generative-ai-guide · GitHub

ii. Mini-Batch Discrimination: This technique encourages the generator to

produce more diverse samples by penalizing mode collapse. By computing
statistics across multiple samples in a mini-batch, the discriminator can
identify mode collapse and provide feedback to the generator to encourage
diversity in the generated samples.
iii. Diverse Training Data: Ensuring that the training dataset contains diverse
samples from the target distribution can help prevent mode collapse. If the
training data is highly skewed or lacks diversity, the generator may struggle
to capture the full complexity of the data distribution.
iv. Regularization Techniques: Techniques such as weight regularization,
dropout, and spectral normalization can be used to regularize the training of
the GAN, making it more resistant to mode collapse. These techniques help
prevent overfitting and encourage the learning of more diverse features.
v. Dynamic Learning Rates: Adjusting the learning rates of the generator and
discriminator dynamically during training can help stabilize the training
process and prevent mode collapse. Techniques such as using learning rate
schedules or adaptive learning rate algorithms can be effective in this
regard.
vi. Ensemble Methods: Training multiple GANs with different initializations or
architectures and combining their outputs using ensemble methods can help
alleviate mode collapse. By leveraging the diversity of multiple generators,
ensemble methods can produce more varied and realistic generated
samples.

6. How does overfitting manifest in generative models, and what techniques can be
used to prevent it during training?
Answer:
Overfitting in generative models occurs when the model memorizes the training
data rather than learning the underlying data distribution, resulting in poor
generalization to new, unseen data. Overfitting can manifest in various ways in
generative models:
i. Mode Collapse: One common manifestation of overfitting in generative
models is mode collapse, where the generator produces a limited variety of
samples, failing to capture the full diversity of the data distribution.
ii. Poor Generalization: Generative models might generate samples that
closely resemble the training data but lack diversity or fail to capture the
nuances present in the true data distribution.
iii. Artifacts or Inconsistencies: Overfitting can lead to the generation of
unrealistic or inconsistent samples, such as distorted images, implausible
[Link] 5/47
17/02/2025, 03:02 awesome-generative-ai-guide/interview_prep/60_gen_ai_questions.md at main · aishwaryanr/awesome-generative-ai-guide · GitHub

text sequences, or nonsensical outputs.

To prevent overfitting in generative models during training, various techniques
can be employed:
i. Regularization: Regularization techniques such as weight decay, dropout,
and batch normalization can help prevent overfitting by imposing constraints
on the model's parameters or introducing stochasticity during training.
ii. Early Stopping: Monitoring the performance of the generative model on a
validation set and stopping training when performance begins to deteriorate
can prevent overfitting and ensure that the model generalizes well to unseen
data.
iii. Data Augmentation: Increasing the diversity of the training data through
techniques like random cropping, rotation, scaling, or adding noise can help
prevent overfitting by exposing the model to a wider range of variations in
the data distribution.
iv. Adversarial Training: Adversarial training, where the generator is trained to
fool a discriminator that is simultaneously trained to distinguish between real
and generated samples, can help prevent mode collapse and encourage the
generation of diverse and realistic samples.
v. Ensemble Methods: Training multiple generative models with different
architectures or initializations and combining their outputs using ensemble
methods can help mitigate overfitting by leveraging the diversity of multiple
models.
vi. Cross-Validation: Partitioning the dataset into multiple folds and training the
model on different subsets while validating on the remaining data can help
prevent overfitting by providing more reliable estimates of the model's
performance on unseen data.

7. What is gradient clipping, and how does it help in stabilizing the training process
of generative models?
Answer:
Gradient clipping is a technique used during training to limit the magnitude of
gradients, typically applied when the gradients exceed a predefined threshold. It
is commonly employed in deep learning models, including generative models like
Generative Adversarial Networks (GANs) and Variational Autoencoders (VAEs).
Gradient clipping helps stabilize the training process of generative models in
several ways:

[Link] 6/47
17/02/2025, 03:02 awesome-generative-ai-guide/interview_prep/60_gen_ai_questions.md at main · aishwaryanr/awesome-generative-ai-guide · GitHub

i. Preventing Exploding Gradients: In deep neural networks, particularly in

architectures with deep layers, gradients can sometimes explode during
training, leading to numerical instability and hindering convergence. Gradient
clipping imposes an upper bound on the gradient values, preventing them
from growing too large and causing numerical issues.
ii. Mitigating Oscillations: During training, gradients can oscillate widely due to
the complex interactions between the generator and discriminator (in GANs)
or the encoder and decoder (in VAEs). Gradient clipping helps dampen these
oscillations by constraining the magnitude of the gradients, leading to
smoother and more stable updates to the model parameters.
iii. Enhancing Convergence: By preventing the gradients from becoming too
large or too small, gradient clipping promotes more consistent and
predictable updates to the model parameters. This can lead to faster
convergence during training, as the model is less likely to encounter extreme
gradient values that impede progress.
iv. Improving Robustness: Gradient clipping can help make the training
process more robust to variations in hyperparameters, such as learning rates
or batch sizes. It provides an additional safeguard against potential
instabilities that may arise due to changes in the training dynamics.

8. Discuss strategies for training generative models when the available dataset is
limited.
Answer:
When dealing with limited datasets, training generative models can be
challenging due to the potential for overfitting and the difficulty of capturing the
full complexity of the underlying data distribution. However, several strategies
can be employed to effectively train generative models with limited data:
i. Data Augmentation: Augmenting the existing dataset by applying
transformations such as rotation, scaling, cropping, or adding noise can
increase the diversity of the training data. This helps prevent overfitting and
enables the model to learn more robust representations of the data
distribution.
ii. Transfer Learning: Leveraging pre-trained models trained on larger datasets
can provide a valuable initialization for the generative model. By fine-tuning
the pre-trained model on the limited dataset, the model can adapt its
representations to the specific characteristics of the target domain more
efficiently.
iii. Semi-supervised Learning: If a small amount of labeled data is available in
addition to the limited dataset, semi-supervised learning techniques can be
[Link] 7/47
17/02/2025, 03:02 awesome-generative-ai-guide/interview_prep/60_gen_ai_questions.md at main · aishwaryanr/awesome-generative-ai-guide · GitHub

employed. These techniques leverage both labeled and unlabeled data to

improve model performance, often by jointly optimizing a supervised and
unsupervised loss function.
iv. Regularization: Regularization techniques such as weight decay, dropout,
and batch normalization can help prevent overfitting by imposing constraints
on the model's parameters or introducing stochasticity during training.
Regularization encourages the model to learn more generalizable
representations of the data.
v. Generative Adversarial Networks (GANs) with Progressive Growing:
Progressive growing GANs (PGGANs) incrementally increase the resolution
of generated images during training, starting from low resolution and
gradually adding detail. This allows the model to learn more effectively from
limited data by focusing on coarse features before refining finer details.
vi. Ensemble Methods: Training multiple generative models with different
architectures or initializations and combining their outputs using ensemble
methods can help mitigate the limitations of a small dataset. Ensemble
methods leverage the diversity of multiple models to improve the overall
performance and robustness of the generative model.
vii. Data Synthesis: In cases where the available dataset is extremely limited,
data synthesis techniques such as generative adversarial networks (GANs)
or variational autoencoders (VAEs) can be used to generate synthetic data
samples. These synthetic samples can be combined with the limited real
data to augment the training dataset and improve model performance.

9. Explain how curriculum learning can be applied in the training of generative

models. What advantages does it offer?
Answer:
Curriculum learning is a training strategy inspired by the way humans learn,
where we often start with simpler concepts and gradually move towards more
complex ones. This approach can be effectively applied in the training of
generative models, a class of AI models designed to generate data similar to
some input data, such as images, text, or sound.

[Link] 8/47
17/02/2025, 03:02 awesome-generative-ai-guide/interview_prep/60_gen_ai_questions.md at main · aishwaryanr/awesome-generative-ai-guide · GitHub

To apply curriculum learning in the training of generative models, you would start
by organizing the training data into a sequence of subsets, ranging from simpler
to more complex examples. The criteria for complexity can vary depending on the
task and the data. For instance, in a text generation task, simpler examples could
be shorter sentences with common vocabulary, while more complex examples
could be longer sentences with intricate structures and diverse vocabulary. In
image generation, simpler examples might include images with less detail or
fewer objects, progressing to more detailed images with complex scenes.
The training process then begins with the model learning from the simpler subset
of data, gradually introducing more complex subsets as the model's performance
improves. This incremental approach helps the model to first grasp basic
patterns before tackling more challenging ones, mimicking a learning progression
that can lead to more efficient and effective learning.
The advantages of applying curriculum learning to the training of generative
models include:
i. Improved Learning Efficiency: Starting with simpler examples can help the
model to quickly learn basic patterns before gradually adapting to more
complex ones, potentially speeding up the training process.
ii. Enhanced Model Performance: By structuring the learning process, the
model may achieve better generalization and performance on complex
examples, as it has built a solid foundation on simpler tasks.
iii. Stabilized Training Process: Gradually increasing the complexity of the
training data can lead to a more stable training process, reducing the risk of
the model getting stuck in poor local minima early in training.
iv. Reduced Overfitting: By effectively learning general patterns from simpler
examples before moving to complex ones, the model might be less prone to
overfitting on the training data.

10. Describe the concept of learning rate scheduling and its role in optimizing the
training process of generative models over time.
Answer:
Learning rate scheduling is a crucial technique in the optimization of neural
networks, including generative models, which involves adjusting the learning rate
—the step size used to update the model's weights—over the course of training.
The learning rate is a critical hyperparameter that determines how much the
model adjusts its weights in response to the estimated error each time it is
updated. If the learning rate is too high, the model may overshoot the optimal
solution; if it's too low, training may proceed very slowly or stall.
[Link] 9/47
17/02/2025, 03:02 awesome-generative-ai-guide/interview_prep/60_gen_ai_questions.md at main · aishwaryanr/awesome-generative-ai-guide · GitHub

In the context of training generative models, such as Generative Adversarial

Networks (GANs) or Variational Autoencoders (VAEs), learning rate scheduling
can significantly impact the model's ability to learn complex data distributions
effectively and efficiently.
Role in Optimizing the Training Process:
i. Avoids Overshooting: Early in training, a higher learning rate can help the
model quickly converge towards a good solution. However, as training
progresses and the model gets closer to the optimal solution, that same high
learning rate can cause the model to overshoot the target. Gradually
reducing the learning rate helps avoid this problem, allowing the model to
fine-tune its parameters more delicately.
ii. Speeds Up Convergence: Initially using a higher learning rate can
accelerate the convergence by allowing larger updates to the weights. This is
especially useful in the early phases of training when the model is far from
the optimal solution.
iii. Improves Model Performance: By carefully adjusting the learning rate over
time, the model can escape suboptimal local minima or saddle points more
efficiently, potentially leading to better overall performance on the
generation task.
iv. Adapts to Training Dynamics: Different phases of training may require
different learning rates. For example, in the case of GANs, the balance
between the generator and discriminator can vary widely during training.
Adaptive learning rate scheduling can help maintain this balance by adjusting
the learning rates according to the training dynamics.
Common Scheduling Strategies:
Step Decay: Reducing the learning rate by a factor every few epochs.
Exponential Decay: Continuously reducing the learning rate exponentially
over time.
Cosine Annealing: Adjusting the learning rate following a cosine function,
leading to periodic adjustments that can help in escaping local minima.
Warm-up Schedules: Gradually increasing the learning rate from a small to
a larger value during the initial phase of training, which can help in stabilizing
the training of very deep models.

11. Compare and contrast the use of L1 and L2 loss functions in the context of
generative models. When might one be preferred over the other?
Answer:
[Link] 10/47
17/02/2025, 03:02 awesome-generative-ai-guide/interview_prep/60_gen_ai_questions.md at main · aishwaryanr/awesome-generative-ai-guide · GitHub

Both loss functions are used to measure the difference between the model's
predictions and the actual data, but they do so in distinct ways that affect the
model's learning behavior and output characteristics.
L1 Loss (Absolute Loss): The L1 loss function calculates the absolute differences
between the predicted values and the actual values. This approach is less
sensitive to outliers because it treats all deviations the same, regardless of their
magnitude. In the context of generative models, using L1 loss can lead to sparser
gradients, which may result in models that are more robust to noise in the input
data. Moreover, L1 loss tends to produce results that are less smooth, which
might be preferable when sharp transitions or details are desired in the
generated outputs, such as in image super-resolution tasks.
L2 Loss (Squared Loss): On the other hand, the L2 loss function computes the
square of the differences between the predicted and actual values. This makes it
more sensitive to outliers, as larger deviations are penalized more heavily. The
use of L2 loss in generative models often results in smoother outcomes because
it encourages smaller and more incremental changes in the model's parameters.
This characteristic can be beneficial in tasks where the continuity of the output is
critical, like generating realistic textures in images.
Preference and Application:
Preference for L1 Loss: You might prefer L1 loss when the goal is to
encourage more robustness to outliers in the dataset or when generating
outputs where precise edges and details are important. Its tendency to
produce sparser solutions can be particularly useful in applications requiring
high detail fidelity, such as in certain types of image processing where
sharpness is key.
Preference for L2 Loss: L2 loss could be the preferred choice when aiming
for smoother outputs and when dealing with problems where the Gaussian
noise assumption is reasonable. Its sensitivity to outliers makes it suitable for
tasks where the emphasis is on minimizing large errors, contributing to
smoother and more continuous generative outputs.

12. In the context of GANs, what is the purpose of gradient penalties in the loss
function? How do they address training instability?
Answer:

[Link] 11/47
17/02/2025, 03:02 awesome-generative-ai-guide/interview_prep/60_gen_ai_questions.md at main · aishwaryanr/awesome-generative-ai-guide · GitHub

Gradient penalties are a crucial technique designed to enhance the stability and
reliability of the training process. GANs consist of two competing networks: a
generator, which creates data resembling the target distribution, and a
discriminator, which tries to distinguish between real data from the target
distribution and fake data produced by the generator. While powerful, GANs are
notorious for their training difficulties, including instability, mode collapse, and
the vanishing gradient problem.
Purpose of Gradient Penalties:
The primary purpose of introducing gradient penalties into the loss function of
GANs is to impose a regularization constraint on the training process. This
constraint ensures that the gradients of the discriminator (with respect to its
input) do not become too large, which is a common source of instability in GAN
training. By penalizing large gradients, these methods encourage smoother
decision boundaries from the discriminator, which, in turn, provides more
meaningful gradients to the generator during backpropagation. This is crucial for
the generator to learn effectively and improve the quality of the generated
samples.
Gradient penalties help to enforce a Lipschitz continuity condition on the
discriminator function. A function is Lipschitz continuous if there exists a
constant such that the function does not change faster than this constant times
the change in input. In the context of GANs, ensuring the discriminator adheres
to this condition helps in stabilizing training by preventing excessively large
updates that can derail the learning process.
Addressing Training Instability:
i. Improved Gradient Flow: By penalizing extreme gradients, gradient
penalties ensure a more stable gradient flow between the discriminator and
the generator. This stability is critical for the generator to learn effectively, as
it relies on feedback from the discriminator to adjust its parameters.
ii. Prevention of Mode Collapse: Mode collapse occurs when the generator
produces a limited variety of outputs. Gradient penalties can mitigate this
issue by ensuring that the discriminator provides consistent and diversified
feedback to the generator, encouraging it to explore a wider range of the
data distribution.
iii. Enhanced Robustness: The regularization effect of gradient penalties
makes the training process more robust to hyperparameter settings and
initialization, reducing the sensitivity of GANs to these factors and making it
easier to achieve convergence.
iv. Encouraging Smooth Decision Boundaries: By enforcing Lipschitz
continuity, gradient penalties encourage the discriminator to form smoother
[Link] 12/47
17/02/2025, 03:02 awesome-generative-ai-guide/interview_prep/60_gen_ai_questions.md at main · aishwaryanr/awesome-generative-ai-guide · GitHub

decision boundaries. This can lead to more gradual transitions in the

discriminator's judgments, providing the generator with more nuanced
gradients for learning to produce high-quality outputs.
Examples of Gradient Penalties:
Wasserstein GAN with Gradient Penalty (WGAN-GP): A well-known variant
that introduces a gradient penalty term to the loss function to enforce the
Lipschitz constraint, significantly improving the stability and quality of the
training process.
Spectral Normalization: Although not a gradient penalty per se, spectral
normalization is another technique to control the Lipschitz constant of the
discriminator by normalizing its weights, which indirectly affects the
gradients and contributes to training stability.

Large Language Models

1. Discuss the concept of transfer learning in the context of natural language
processing. How do pre-trained language models contribute to various NLP
tasks?
Answer:
Transfer learning typically involves two main phases:
i. Pre-training: In this phase, a language model is trained on a large corpus of
text data. This training is unsupervised or semi-supervised and aims to learn
a general understanding of the language, including its syntax, semantics,
and context. Models learn to predict the next word in a sentence, fill in
missing words, or even predict words based on their context in a
bidirectional manner.
ii. Fine-tuning: After the pre-training phase, the model is then fine-tuned on a
smaller, task-specific dataset. During fine-tuning, the model's parameters
are slightly adjusted to specialize in the specific NLP task at hand, such as
sentiment analysis, question-answering, or text classification. The idea is
that the model retains its general understanding of the language learned
during pre-training while adapting to the nuances of the specific task.
Pre-trained language models have revolutionized NLP by providing a strong
foundational knowledge of language that can be applied to a multitude of tasks.
Some key contributions include:
Improved Performance: Pre-trained models have set new benchmarks
across various NLP tasks by leveraging their extensive pre-training on
[Link] 13/47
17/02/2025, 03:02 awesome-generative-ai-guide/interview_prep/60_gen_ai_questions.md at main · aishwaryanr/awesome-generative-ai-guide · GitHub

diverse language data. This has led to significant improvements in tasks

such as text classification, named entity recognition, machine translation,
and more.
Efficiency in Training: By starting with a model that already understands
language to a significant degree, researchers and practitioners can achieve
high performance on specific tasks with relatively little task-specific data.
This drastically reduces the resources and time required to train models from
scratch.
Versatility: The same pre-trained model can be fine-tuned for a wide range
of tasks without substantial modifications. This versatility makes pre-trained
language models highly valuable across different domains and applications,
from healthcare to legal analysis.
Handling of Contextual Information: Models like BERT (Bidirectional
Encoder Representations from Transformers) and its successors (e.g.,
RoBERTa, GPT-3) are particularly adept at understanding the context of
words in a sentence, leading to more nuanced and accurate interpretations
of text. This capability is crucial for complex tasks such as sentiment
analysis, where the meaning can significantly depend on context.
Language Understanding: Pre-trained models have advanced the
understanding of language nuances, idioms, and complex sentence
structures. This has improved machine translation and other tasks requiring
deep linguistic insights.

2. Highlight the key differences between models like GPT (Generative Pre-trained
Transformer) and BERT (Bidirectional Encoder Representations from
Transformers)?
Answer:

[Link] 14/47
17/02/2025, 03:02 awesome-generative-ai-guide/interview_prep/60_gen_ai_questions.md at main · aishwaryanr/awesome-generative-ai-guide · GitHub

Image Source: [Link]

and-gpt/
GPT (Generative Pre-trained Transformer) and BERT (Bidirectional Encoder
Representations from Transformers) are two foundational architectures in the
field of NLP (Natural Language Processing), each with its unique approach and
capabilities. Although both models leverage the Transformer architecture for
processing text, they are designed for different purposes and operate in distinct
ways.
Architecture and Training Approach:
GPT:
GPT is designed as an autoregressive model that predicts the next word
in a sequence given the previous words. Its training is based on the left-
to-right context only.
It is primarily used for generative tasks, where the model generates text
based on the input it receives.
GPT's architecture is a stack of Transformer decoder blocks.
BERT:
BERT, in contrast, is designed to understand the context of words in a
sentence by considering both left and right contexts (i.e.,
bidirectionally). It does not predict the next word in a sequence but
rather learns word representations that reflect both preceding and
following words.
BERT is pre-trained using two strategies: Masked Language Model
(MLM) and Next Sentence Prediction (NSP). MLM involves randomly
masking words in a sentence and then predicting them based on their
[Link] 15/47
17/02/2025, 03:02 awesome-generative-ai-guide/interview_prep/60_gen_ai_questions.md at main · aishwaryanr/awesome-generative-ai-guide · GitHub

context, while NSP involves predicting whether two sentences logically

follow each other.
BERT's architecture is a stack of Transformer encoder blocks.
Use Cases and Applications:
GPT:
Given its generative nature, GPT excels in tasks that require content
generation, such as creating text, code, or even poetry. It is also
effective in tasks like language translation, text summarization, and
question-answering where generating coherent and contextually
relevant text is crucial.
BERT:
BERT is particularly effective for tasks that require understanding the
context and nuances of language, such as sentiment analysis, named
entity recognition (NER), and question answering where the model
provides answers based on given content rather than generating new
content.
Training and Fine-tuning:
GPT:
GPT models are trained on a large corpus of text in an unsupervised
manner and then fine-tuned for specific tasks by adjusting the model on
a smaller, task-specific dataset.
BERT:
BERT is also pre-trained on a large text corpus but uses a different set
of pre-training objectives. Its fine-tuning process is similar to GPT's,
where the pre-trained model is adapted to specific tasks with additional
task-specific layers if necessary.
Performance and Efficiency:
GPT:
GPT models, especially in their later iterations like GPT-3, have shown
remarkable performance in generating human-like text. However, their
autoregressive nature can sometimes lead to less efficiency in tasks that
require understanding the full context of input text.
BERT:
BERT has been a breakthrough in tasks requiring deep understanding of
context and relationships within text. Its bidirectional nature allows it to
outperform or complement autoregressive models in many such tasks.

[Link] 16/47
17/02/2025, 03:02 awesome-generative-ai-guide/interview_prep/60_gen_ai_questions.md at main · aishwaryanr/awesome-generative-ai-guide · GitHub

3. What problems of RNNs do transformer models solve?

Answer:
Transformer models were designed to overcome several significant limitations
associated with Recurrent Neural Networks, including:
Difficulty with Parallelization: RNNs process data sequentially, which
inherently limits the possibility of parallelizing computations. Transformers,
by contrast, leverage self-attention mechanisms to process entire
sequences simultaneously, drastically improving efficiency and reducing
training time.
Long-Term Dependencies: RNNs, especially in their basic forms, struggle
with capturing long-term dependencies due to vanishing and exploding
gradient problems. Transformers address this by using self-attention
mechanisms that directly compute relationships between all parts of the
input sequence, regardless of their distance from each other.
Scalability: The sequential nature of RNNs also makes them less scalable for
processing long sequences, as computational complexity and memory
requirements increase linearly with sequence length. Transformers mitigate
this issue through more efficient attention mechanisms, although they still
face challenges with very long sequences without modifications like sparse
attention patterns.

4. Why is incorporating relative positional information crucial in transformer

models? Discuss scenarios where relative position encoding is particularly
beneficial.
Answer:
In transformer models, understanding the sequence's order is essential since the
self-attention mechanism treats each input independently of its position in the
sequence. Incorporating relative positional information allows transformers to
capture the order and proximity of elements, which is crucial for tasks where the
meaning depends significantly on the arrangement of components.
Relative position encoding is particularly beneficial in:
Language Understanding and Generation: The meaning of a sentence can
change dramatically based on word order. For example, "The cat chased the
mouse" versus "The mouse chased the cat."

[Link] 17/47
17/02/2025, 03:02 awesome-generative-ai-guide/interview_prep/60_gen_ai_questions.md at main · aishwaryanr/awesome-generative-ai-guide · GitHub

Sequence-to-Sequence Tasks: In machine translation, maintaining the correct

order of words is vital for accurate translations. Similarly, for tasks like text
summarization, understanding the relative positions helps in identifying key
points and their significance within the text.
Time-Series Analysis: When transformers are applied to time-series data, the
relative positioning helps the model understand temporal relationships, such as
causality and trends over time.

5. What challenges arise from the fixed and limited attention span in the vanilla
Transformer model? How does this limitation affect the model's ability to capture
long-term dependencies?
Answer
The vanilla Transformer model has a fixed attention span, typically limited by the
maximum sequence length it can process, which poses challenges in capturing
long-term dependencies in extensive texts. This limitation stems from the
quadratic complexity of the self-attention mechanism with respect to sequence
length, leading to increased computational and memory requirements for longer
sequences.
This limitation affects the model's ability in several ways:
Difficulty in Processing Long Documents: For tasks such as document
summarization or long-form question answering, the model may struggle to
integrate critical information spread across a large document.
Impaired Contextual Understanding: In narrative texts or dialogues where
context from early parts influences the meaning of later parts, the model's fixed
attention span may prevent it from fully understanding or generating coherent
and contextually consistent text.

6. Why is naively increasing context length not a straightforward solution for

handling longer context in transformer models? What computational and memory
challenges does it pose?
Answer:
Naively increasing the context length in transformer models to handle longer
contexts is not straightforward due to the self-attention mechanism's quadratic
computational and memory complexity with respect to sequence length. This
increase in complexity means that doubling the sequence length quadruples the
computation and memory needed, leading to:
[Link] 18/47
17/02/2025, 03:02 awesome-generative-ai-guide/interview_prep/60_gen_ai_questions.md at main · aishwaryanr/awesome-generative-ai-guide · GitHub

Excessive Computational Costs: Processing longer sequences requires

significantly more computing power, slowing down both training and inference
times.
Memory Constraints: The increased memory demand can exceed the capacity
of available hardware, especially GPUs, limiting the feasibility of processing long
sequences and scaling models effectively.

7. How does self-attention work?

Answer:
Self-attention is a mechanism that enables models to weigh the importance of
different parts of the input data relative to each other. In the context of
transformers, it allows every output element to be computed as a weighted sum
of a function of all input elements, enabling the model to focus on different parts
of the input sequence when performing a specific task. The self-attention
mechanism involves three main steps:
Query, Key, and Value Vectors: For each input element, the model generates
three vectors—a query vector, a key vector, and a value vector—using learnable
weights.
Attention Scores: The model calculates attention scores by performing a dot
product between the query vector of one element and the key vector of every
other element, followed by a softmax operation to normalize the scores. These
scores determine how much focus or "attention" each element should give to
every other element in the sequence.
Weighted Sum and Output: The attention scores are used to create a weighted
sum of the value vectors, which forms the output for each element. This process
allows the model to dynamically prioritize information from different parts of the
input sequence based on the

[Link] 19/47
17/02/2025, 03:02 awesome-generative-ai-guide/interview_prep/60_gen_ai_questions.md at main · aishwaryanr/awesome-generative-ai-guide · GitHub

8. What pre-training mechanisms are used for LLMs, explain a few

Answer**:**
Large Language Models utilize several pre-training mechanisms to learn from
vast amounts of text data before being fine-tuned on specific tasks. Key
mechanisms include:
Masked Language Modeling (MLM): Popularized by BERT, this involves
randomly masking some percentage of the input tokens and training the model to
predict these masked tokens based on their context. This helps the model learn a
deep understanding of language context and structure.
Causal (Autoregressive) Language Modeling: Used by models like GPT, this
approach trains the model to predict the next token in a sequence based on the
tokens that precede it. This method is particularly effective for generative tasks
where the model needs to produce coherent and contextually relevant text.

[Link] 20/47
17/02/2025, 03:02 awesome-generative-ai-guide/interview_prep/60_gen_ai_questions.md at main · aishwaryanr/awesome-generative-ai-guide · GitHub

Permutation Language Modeling: Introduced by XLNet, this technique involves

training the model to predict a token within a sequence given the other tokens,
where the order of the input tokens is permuted. This encourages the model to
understand language in a more flexible and context-aware manner.

9. Why is a multi-head attention needed?

Answer:
Answer: Multi-head attention allows a model to jointly attend to information from
different representation subspaces at different positions. This is achieved by
running several attention mechanisms (heads) in parallel, each with its own set of
learnable weights. The key benefits include:
Richer Representation: By capturing different aspects of the information (e.g.,
syntactic and semantic features) in parallel, multi-head attention allows the
model to develop a more nuanced understanding of the input.
Improved Attention Focus: Different heads can focus on different parts of the
sequence, enabling the model to balance local and global information and
improve its ability to capture complex dependencies.
Increased Model Capacity: Without significantly increasing computational
complexity, multi-head attention provides a way to increase the model's capacity,
allowing it to learn more complex patterns and relationships in the data.

10. What is RLHF, how is it used?

Answer:

[Link] 21/47
17/02/2025, 03:02 awesome-generative-ai-guide/interview_prep/60_gen_ai_questions.md at main · aishwaryanr/awesome-generative-ai-guide · GitHub

Image Source: [Link]

Reinforcement Learning from Human Feedback (RLHF) is a method used to fine-
tune language models in a way that aligns their outputs with human preferences,
values, and ethics. The process involves several steps:
Pre-training: The model is initially pre-trained on a large corpus of text data to
learn a broad understanding of language.
Human Feedback Collection: Human annotators review the model's outputs in
specific scenarios and provide feedback or corrections.
Reinforcement Learning: The model is fine-tuned using reinforcement learning
techniques, where the human feedback serves as a reward signal, encouraging
the model to produce outputs that are more aligned with human judgments.
RLHF is particularly useful for tasks requiring a high degree of alignment with
human values, such as generating safe and unbiased content, enhancing the
quality of conversational agents, or ensuring that AI-generated advice is ethically
sound.
Read the article form Huggingface: [Link]

[Link] 22/47
17/02/2025, 03:02 awesome-generative-ai-guide/interview_prep/60_gen_ai_questions.md at main · aishwaryanr/awesome-generative-ai-guide · GitHub

11. What is catastrophic forgetting in the context of LLMs

Answer:
Catastrophic forgetting refers to the phenomenon where a neural network,
including Large Language Models, forgets previously learned information upon
learning new information. This occurs because neural networks adjust their
weights during training to minimize the loss on the new data, which can
inadvertently cause them to "forget" what they had learned from earlier data.
This issue is particularly challenging in scenarios where models need to
continuously learn from new data streams without losing their performance on
older tasks.

12. In a transformer-based sequence-to-sequence model, what are the primary

functions of the encoder and decoder? How does information flow between them
during both training and inference?
Answer:
In a transformer-based sequence-to-sequence model, the encoder and decoder
serve distinct but complementary roles in processing and generating sequences:
Encoder: The encoder processes the input sequence, capturing its informational
content and contextual relationships. It transforms the input into a set of
continuous representations, which encapsulate the input sequence's information
in a form that the decoder can utilize.
Decoder: The decoder receives the encoder's output representations and
generates the output sequence, one element at a time. It uses the encoder's
representations along with the previously generated elements to produce the
next element in the sequence.
During training and inference, information flows between the encoder and
decoder primarily through the encoder's output representations. In addition, the
decoder uses self-attention to consider its previous outputs when generating the
next output, ensuring coherence and contextuality in the generated sequence. In
some transformer variants, cross-attention mechanisms in the decoder also allow
direct attention to the encoder's outputs at each decoding step, further
enhancing the model's ability to generate relevant and accurate sequences
based on the input.

13. Why is positional encoding crucial in transformer models, and what issue does it
address in the context of self-attention operations?
[Link] 23/47
17/02/2025, 03:02 awesome-generative-ai-guide/interview_prep/60_gen_ai_questions.md at main · aishwaryanr/awesome-generative-ai-guide · GitHub

Answer:
Positional encoding is a fundamental aspect of transformer models, designed to
imbue them with the ability to recognize the order of elements in a sequence.
This capability is crucial because the self-attention mechanism at the heart of
transformer models treats each element of the input sequence independently,
without any inherent understanding of the position or order of elements. Without
positional encoding, transformers would not be able to distinguish between
sequences of the same set of elements arranged in different orders, leading to a
significant loss in the ability to understand and generate meaningful language or
process sequence data effectively.
Addressing the Issue of Sequence Order in Self-Attention Operations:
The self-attention mechanism allows each element in the input sequence to
attend to all elements simultaneously, calculating the attention scores based on
the similarity of their features. While this enables the model to capture complex
relationships within the data, it inherently lacks the ability to understand how the
position of an element in the sequence affects its meaning or role. For example,
in language, the meaning of a sentence can drastically change with the order of
words ("The cat ate the fish" vs. "The fish ate the cat"), and in time-series data,
the position of data points in time is critical to interpreting patterns and trends.
How Positional Encoding Works:
To overcome this limitation, positional encodings are added to the input
embeddings at the beginning of the transformer model. These encodings provide
a unique signature for each position in the sequence, which is combined with the
element embeddings, thus allowing the model to retain and utilize positional
information throughout the self-attention and subsequent layers. Positional
encodings can be designed in various ways, but they typically involve patterns
that the model can learn to associate with sequence order, such as sinusoidal
functions of different frequencies.

14. When applying transfer learning to fine-tune a pre-trained transformer for a

specific NLP task, what strategies can be employed to ensure effective
knowledge transfer, especially when dealing with domain-specific data?
Answer:
Applying transfer learning to fine-tune a pre-trained transformer model involves
several strategies to ensure that the vast knowledge the model has acquired is
effectively transferred to the specific requirements of a new, potentially domain-
specific task:
[Link] 24/47
17/02/2025, 03:02 awesome-generative-ai-guide/interview_prep/60_gen_ai_questions.md at main · aishwaryanr/awesome-generative-ai-guide · GitHub

Domain-Specific Pre-training: Before fine-tuning on the task-specific dataset,

pre-train the model further on a large corpus of domain-specific data. This step
helps the model to adapt its general language understanding capabilities to the
nuances, vocabulary, and stylistic features unique to the domain in question.
Gradual Unfreezing: Start fine-tuning by only updating the weights of the last
few layers of the model and gradually unfreeze more layers as training
progresses. This approach helps in preventing the catastrophic forgetting of pre-
trained knowledge while allowing the model to adapt to the specifics of the new
task.
Learning Rate Scheduling: Employ differential learning rates across the layers of
the model during fine-tuning. Use smaller learning rates for earlier layers, which
contain more general knowledge, and higher rates for later layers, which are
more task-specific. This strategy balances retaining what the model has learned
with adapting to new data.
Task-Specific Architectural Adjustments: Depending on the task, modify the
model architecture by adding task-specific layers or heads. For instance, adding
a classification head for a sentiment analysis task or a sequence generation head
for a translation task allows the model to better align its outputs with the
requirements of the task.
Data Augmentation: Increase the diversity of the task-specific training data
through techniques such as back-translation, synonym replacement, or sentence
paraphrasing. This can help the model generalize better across the domain-
specific nuances.
Regularization Techniques: Implement techniques like dropout, label smoothing,
or weight decay during fine-tuning to prevent overfitting to the smaller, task-
specific dataset, ensuring the model retains its generalizability.

15. Discuss the role of cross-attention in transformer-based encoder-decoder

models. How does it facilitate the generation of output sequences based on
information from the input sequence?
Answer:
Cross-attention is a mechanism in transformer-based encoder-decoder models
that allows the decoder to focus on different parts of the input sequence as it
generates each token of the output sequence. It plays a crucial role in tasks such
as machine translation, summarization, and question answering, where the
output depends directly on the input content.

[Link] 25/47
17/02/2025, 03:02 awesome-generative-ai-guide/interview_prep/60_gen_ai_questions.md at main · aishwaryanr/awesome-generative-ai-guide · GitHub

During the decoding phase, for each output token being generated, the cross-
attention mechanism queries the encoder's output representations with the
current state of the decoder. This process enables the decoder to "attend" to the
most relevant parts of the input sequence, extracting the necessary information
to generate the next token in the output sequence. Cross-attention thus
facilitates a dynamic, content-aware generation process where the focus shifts
across different input elements based on their relevance to the current decoding
step.
This ability to selectively draw information from the input sequence ensures that
the generated output is contextually aligned with the input, enhancing the
coherence, accuracy, and relevance of the generated text.

16. ****Compare and contrast the impact of using sparse (e.g., cross-entropy) and
dense (e.g., mean squared error) loss functions in training language models.
Answer:
Sparse and dense loss functions serve different roles in the training of language
models, impacting the learning process and outcomes in distinct ways:
Sparse Loss Functions (e.g., Cross-Entropy): These are typically used in
classification tasks, including language modeling, where the goal is to predict the
next word from a large vocabulary. Cross-entropy measures the difference
between the predicted probability distribution over the vocabulary and the actual
distribution (where the actual word has a probability of 1, and all others are 0). It
is effective for language models because it directly penalizes the model for
assigning low probabilities to the correct words and encourages sparsity in the
output distribution, reflecting the reality that only a few words are likely at any
given point.
Dense Loss Functions (e.g., Mean Squared Error (MSE)): MSE measures the
average of the squares of the differences between predicted and actual values.
While not commonly used for categorical outcomes like word predictions in
language models, it is more suited to regression tasks. In the context of
embedding-based models or continuous output tasks within NLP, dense loss
functions could be applied to measure how closely the generated embeddings
match expected embeddings.
Impact on Training and Model Performance:

[Link] 26/47
17/02/2025, 03:02 awesome-generative-ai-guide/interview_prep/60_gen_ai_questions.md at main · aishwaryanr/awesome-generative-ai-guide · GitHub

Focus on Probability Distribution: Sparse loss functions like cross-entropy align

well with the probabilistic nature of language, focusing on improving the
accuracy of probability distribution predictions for the next word. They are
particularly effective for discrete output spaces, such as word vocabularies in
language models.
Sensitivity to Output Distribution: Dense loss functions, when applied in
relevant NLP tasks, would focus more on minimizing the average error across all
outputs, which can be beneficial for tasks involving continuous data or
embeddings. However, they might not be as effective for typical language
generation tasks due to the categorical nature of text.

17. How can reinforcement learning be integrated into the training of large language
models, and what challenges might arise in selecting suitable loss functions for
RL-based approaches?
Answer:
Integrating reinforcement learning (RL) into the training of large language models
involves using reward signals to guide the model's generation process towards
desired outcomes. This approach, often referred to as Reinforcement Learning
from Human Feedback (RLHF), can be particularly effective for tasks where
traditional supervised learning methods fall short, such as ensuring the
generation of ethical, unbiased, or stylistically specific text.
Integration Process:
Reward Modeling: First, a reward model is trained to predict the quality of model
outputs based on criteria relevant to the task (e.g., coherence, relevance, ethics).
This model is typically trained on examples rated by human annotators.
Policy Optimization: The language model (acting as the policy in RL
terminology) is then fine-tuned using gradients estimated from the reward
model, encouraging the generation of outputs that maximize the predicted
rewards.
Challenges in Selecting Suitable Loss Functions:
Defining Reward Functions: One of the primary challenges is designing or
selecting a reward function that accurately captures the desired outcomes of the
generation task. The reward function must be comprehensive enough to guide
the model towards generating high-quality, task-aligned content without
unintended biases or undesirable behaviors.

[Link] 27/47
17/02/2025, 03:02 awesome-generative-ai-guide/interview_prep/60_gen_ai_questions.md at main · aishwaryanr/awesome-generative-ai-guide · GitHub

Variance and Stability: RL-based approaches can introduce high variance and
instability into the training process, partly due to the challenge of estimating
accurate gradients based on sparse or delayed rewards. Selecting or designing
loss functions that can mitigate these issues is crucial for successful integration.
Reward Shaping and Alignment: Ensuring that the reward signals align with
long-term goals rather than encouraging short-term, superficial optimization is
another challenge. This requires careful consideration of how rewards are
structured and potentially the use of techniques like reward shaping or
constrained optimization.
Integrating RL into the training of large language models holds the promise of
more nuanced and goal-aligned text generation capabilities. However, it requires
careful design and implementation of reward functions and loss calculations to
overcome the inherent challenges of applying RL in complex, high-dimensional
spaces like natural language.

Multimodal Models (Includes non-generative models)

1. In multimodal language models, how is information from visual and textual
modalities effectively integrated to perform tasks such as image captioning or visual
question answering?
Answer:
Multimodal language models integrate visual and textual information through
sophisticated architectures that allow for the processing and analysis of data
from both modalities. These models typically utilize a combination of
convolutional neural networks (CNNs) for image processing and transformers or
recurrent neural networks (RNNs) for text processing. The integration of
information occurs in several ways:
Joint Embedding Space: Both visual and textual inputs are mapped to a
common embedding space where their representations can be compared
directly. This allows the model to understand and manipulate both types of
information in a unified manner.
Attention Mechanisms: Attention mechanisms, particularly cross-modal
attention, enable the model to focus on specific parts of an image given a textual
query (or vice versa), facilitating detailed analysis and understanding of the
relationships between visual and textual elements.

[Link] 28/47
17/02/2025, 03:02 awesome-generative-ai-guide/interview_prep/60_gen_ai_questions.md at main · aishwaryanr/awesome-generative-ai-guide · GitHub

Fusion Layers: After initial processing, the features from both modalities are
combined using fusion layers, which might involve concatenation, element-wise
addition, or more complex interactions. This fusion allows the model to leverage
combined information for tasks like image captioning, where the model generates
descriptive text for an image, or visual question answering, where the model
answers questions based on the content of an image.

2. Explain the role of cross-modal attention mechanisms in models like VisualBERT or

CLIP. How do these mechanisms enable the model to capture relationships between
visual and textual elements?
Answer:
Cross-modal attention mechanisms are pivotal in models like VisualBERT and
CLIP, enabling these systems to dynamically focus on relevant parts of visual
data in response to textual cues and vice versa. This mechanism works by
allowing one modality (e.g., text) to guide the attention process in the other
modality (e.g., image), thereby highlighting the features or areas that are most
relevant to the task at hand.

VisualBERT: Uses cross-modal attention within the transformer architecture to

attend to specific regions of an image based on the context of the text. This is
crucial for tasks where understanding the visual context is essential for
interpreting the textual content correctly.

[Link] 29/47
17/02/2025, 03:02 awesome-generative-ai-guide/interview_prep/60_gen_ai_questions.md at main · aishwaryanr/awesome-generative-ai-guide · GitHub

CLIP: Though not using cross-modal attention in the same way as VisualBERT,
CLIP learns to associate images and texts effectively by training on a vast dataset
of image-text pairs. It uses contrastive learning to maximize the similarity
between corresponding text and image embeddings while minimizing the
similarity between non-corresponding pairs.
In both cases, the cross-modal attention or learning mechanisms allow the
models to understand and leverage the complex relationships between visual
elements and textual descriptions, improving their performance on tasks that
require a nuanced understanding of both modalities.

3. For tasks like image-text matching, how is the training data typically annotated to
create aligned pairs of visual and textual information, and what considerations should
be taken into account?
Answer:
For image-text matching tasks, the training data consists of pairs of images and
textual descriptions that are closely aligned in terms of content and context.
Annotating such data typically involves:

[Link] 30/47
17/02/2025, 03:02 awesome-generative-ai-guide/interview_prep/60_gen_ai_questions.md at main · aishwaryanr/awesome-generative-ai-guide · GitHub

Manual Annotation: Human annotators describe images or annotate existing

descriptions to ensure they accurately reflect the visual content. This process
requires careful guideline development to maintain consistency and accuracy in
the descriptions.
Automated Techniques: Some datasets are compiled using automated
techniques, such as scraping image-caption pairs from the web. However, these
methods require subsequent cleaning and verification to ensure high data quality.
Considerations: When annotating data, it's important to consider diversity (in
terms of both imagery and language), bias (to avoid reinforcing stereotypes or
excluding groups), and specificity (descriptions should be detailed and closely
aligned with the visual content). Additionally, the scalability of the annotation
process is a practical concern, especially for large datasets.

4. When training a generative model for image synthesis, what are common loss
functions used to evaluate the difference between generated and target images, and
how do they contribute to the training process?
Answer:
In image synthesis, common loss functions include:
Pixel-wise Loss Functions: Such as Mean Squared Error (MSE) or Mean
Absolute Error (MAE), which measure the difference between corresponding
pixels in the generated and target images. These loss functions are
straightforward and contribute to ensuring overall fidelity but may not capture
perceptual similarities well.
Adversarial Loss: Used in Generative Adversarial Networks (GANs), where a
discriminator model is trained to distinguish between real and generated images,
providing a signal to the generator on how to improve. This loss function
encourages the generation of images that are indistinguishable from real images,
contributing to the realism of synthesized images.
Perceptual Loss: Measures the difference in high-level features extracted from
pre-trained deep neural networks. This loss function is designed to capture
perceptual and semantic similarities between images, contributing to the
generation of visually and contextually coherent images.

5. What is perceptual loss, and how is it utilized in image generation tasks to measure
the perceptual similarity between generated and target images? How does it differ
from traditional pixel-wise loss functions?
[Link] 31/47
17/02/2025, 03:02 awesome-generative-ai-guide/interview_prep/60_gen_ai_questions.md at main · aishwaryanr/awesome-generative-ai-guide · GitHub

Answer:
Perceptual loss measures the difference in high-level features between the
generated and target images, as extracted by a pre-trained deep neural network
(usually a CNN trained on a large image classification task). This approach
focuses on perceptual and semantic similarities rather than pixel-level accuracy.
Utilization in Image Generation: Perceptual loss is used to guide the training of
generative models by encouraging them to produce images that are similar to the
target images in terms of content and style, rather than exactly matching pixel
values. This is particularly useful for tasks like style transfer, super-resolution,
and photorealistic image synthesis, where the goal is to generate images that
look visually pleasing and coherent to human observers.
Difference from Pixel-wise Loss Functions: Unlike pixel-wise loss functions
(e.g., MSE or MAE) that measure the direct difference between corresponding
pixels, perceptual loss operates at a higher level of abstraction, capturing
differences in textures, shapes, and patterns that contribute to the overall
perception of the image. This makes it more aligned with human visual
perception, leading to more aesthetically pleasing and contextually appropriate
image synthesis.

6. What is Masked language-image modeling?

Answer:
Masked language-image modeling is a training technique used in multimodal
models to learn joint representations of textual and visual information. Similar to
the masked language modeling approach used in BERT for text, this method
involves randomly masking out parts of the input (both in the image and the text)
and training the model to predict the masked elements based on the context
provided by the unmasked elements.
In Images: This might involve masking portions of the image and asking the
model to predict the missing content based on the surrounding visual context
and any associated text.
In Text: Similarly, words or phrases in the text may be masked, and the model
must use the visual context along with the remaining text to predict the missing
words.
This approach encourages the model to develop a deep, integrated
understanding of the content and context across both modalities, enhancing its
capabilities in tasks that require nuanced understanding and manipulation of
visual and textual information.
[Link] 32/47
17/02/2025, 03:02 awesome-generative-ai-guide/interview_prep/60_gen_ai_questions.md at main · aishwaryanr/awesome-generative-ai-guide · GitHub

7. How do attention weights obtained from the cross-attention mechanism

influence the generation process in multimodal models? What role do these
weights play in determining the importance of different modalities?
Answer:
In multimodal models, attention weights obtained from the cross-attention
mechanism play a crucial role in the generation process by dynamically
determining how much importance to give to different parts of the input from
different modalities. These weights influence the model's focus during the
generation process in several ways:
Highlighting Relevant Information: The attention weights enable the model to
focus on the most relevant parts of the visual input when processing textual
information and vice versa. For example, when generating a caption for an image,
the model can focus on specific regions of the image that are most pertinent to
the words being generated.
Balancing Modalities: The weights help in balancing the influence of each
modality on the generation process. Depending on the task and the context, the
model might rely more heavily on textual information in some instances and on
visual information in others. The attention mechanism dynamically adjusts this
balance.
Enhancing Contextual Understanding: By allowing the model to draw on context
from both modalities, the attention weights contribute to a richer, more nuanced
understanding of the input, leading to more accurate and contextually
appropriate outputs.
The ability of cross-attention mechanisms to modulate the influence of different
modalities through attention weights is a powerful feature of multimodal models,
enabling them to perform complex tasks that require an integrated understanding
of visual and textual information.

8. What are the unique challenges in training multimodal generative models

compared to unimodal generative models?
Answer:
Training multimodal generative models introduces unique challenges not typically
encountered in unimodal generative models:

[Link] 33/47
17/02/2025, 03:02 awesome-generative-ai-guide/interview_prep/60_gen_ai_questions.md at main · aishwaryanr/awesome-generative-ai-guide · GitHub

Data Alignment: One of the primary challenges is ensuring proper alignment

between different modalities. For instance, matching specific parts of an image
with corresponding textual descriptions requires sophisticated modeling
techniques to accurately capture and reflect these relationships.
Complexity and Scalability: Multimodal generative models deal with data of
different types (e.g., text, images, audio), each requiring different processing
pipelines. Managing this complexity while scaling the model to handle large
datasets effectively is a significant challenge.
Cross-Modal Coherence: Generating coherent output that makes sense across
all modalities (e.g., an image that accurately reflects a given text description) is
challenging. The model must understand and maintain the context and semantics
across modalities.
Diverse Data Representation: Different modalities have inherently different data
representations (e.g., pixels for images, tokens for text). Designing a model
architecture that can handle these diverse representations and still learn
meaningful cross-modal interactions is challenging.
Sparse Data: In many cases, comprehensive datasets that cover the vast
spectrum of possible combinations of modalities are not available, leading to
sparse data issues. This can make it difficult for the model to learn certain cross-
modal relationships.

9. How do multimodal generative models address the issue of data sparsity in

training?
Answer:
Current multimodal generative models employ several strategies to mitigate the
issue of data sparsity during training:
Data Augmentation: By artificially augmenting the dataset (e.g., generating new
image-text pairs through transformations or translations), models can be
exposed to a broader range of examples, helping to fill gaps in the training data.
Transfer Learning: Leveraging pre-trained models on large unimodal datasets
can provide a strong foundational knowledge that the multimodal model can build
upon. This approach helps the model to generalize better across sparse
multimodal datasets.
Few-Shot and Zero-Shot Learning: These techniques are particularly useful for
handling data sparsity by enabling models to generalize to new, unseen examples
with minimal or no additional training data.
[Link] 34/47
17/02/2025, 03:02 awesome-generative-ai-guide/interview_prep/60_gen_ai_questions.md at main · aishwaryanr/awesome-generative-ai-guide · GitHub

Synthetic Data Generation: Generating synthetic examples of underrepresented

modalities or combinations can help to balance the dataset and provide more
comprehensive coverage of the possible input space.
Regularization Techniques: Implementing regularization methods can prevent
overfitting on the limited available data, helping the model to better generalize
across sparse examples.

10. Explain the concept of Vision-Language Pre-training (VLP) and its significance in
developing robust vision-language models.
Answer:
Vision-Language Pre-training involves training models on large datasets
containing both visual (images, videos) and textual data to learn general
representations that can be fine-tuned for specific vision-language tasks. VLP is
significant because it allows models to capture rich, cross-modal semantic
relationships between visual and textual information, leading to improved
performance on tasks like visual question answering, image captioning, and text-
based image retrieval. By leveraging pre-trained VLP models, developers can
achieve state-of-the-art results on various vision-language tasks with relatively
smaller datasets during fine-tuning, enhancing the model's understanding and
processing of multimodal information.

11. How do models like CLIP and DALL-E demonstrate the integration of vision and
language modalities?
Answer:
CLIP (Contrastive Language-Image Pre-training) and DALL-E (a model designed
for generating images from textual descriptions) are two prominent examples of
models that integrate vision and language modalities effectively:
CLIP: CLIP learns visual concepts from natural language descriptions, training on
a diverse range of images paired with textual descriptions. It uses a contrastive
learning approach to align the image and text representations in a shared
embedding space, enabling it to perform a wide range of vision tasks using
natural language as input. CLIP demonstrates the power of learning from natural
language supervision and its ability to generalize across different vision tasks
without task-specific training data.

[Link] 35/47
17/02/2025, 03:02 awesome-generative-ai-guide/interview_prep/60_gen_ai_questions.md at main · aishwaryanr/awesome-generative-ai-guide · GitHub

DALL-E: DALL-E generates images from textual descriptions, demonstrating a

deep understanding of both the content described in the text and how that
content is visually represented. It uses a version of the GPT-3 architecture
adapted for generating images, showcasing the integration of vision and
language by creating coherent and often surprisingly accurate visual
representations of described scenes, objects, and concepts.
These models exemplify the potential of vision-language integration, highlighting
how deep learning can bridge the gap between textual descriptions and visual
representations to enable creative and flexible applications.

12. How do attention mechanisms enhance the performance of vision-language

models?
Answer:
Attention mechanisms significantly enhance the performance of vision-language
models in multimodal learning by allowing models to dynamically focus on
relevant parts of the input data:
Cross-Modal Attention: These mechanisms enable the model to attend to
specific regions of an image given textual input or vice versa. This selective
attention helps the model to extract and integrate relevant information from both
modalities, improving its ability to perform tasks such as image captioning or
visual question answering by focusing on the salient details that are most
pertinent to the task at hand.
Self-Attention in Language: Within the language modality, self-attention allows
the model to emphasize important words or phrases in a sentence, aiding in
understanding textual context and semantics that are relevant to the visual data.
Self-Attention in Vision: In the visual modality, self-attention mechanisms can
highlight important areas or features within an image, helping to better align
these features with textual descriptions or queries.
By leveraging attention mechanisms, vision-language models can achieve a more
nuanced and effective integration of information across modalities, leading to
more accurate, context-aware, and coherent multimodal representations and
outputs.

Embeddings
1. What is the fundamental concept of embeddings in machine learning, and how do
they represent information in a more compact form compared to raw input data?
[Link] 36/47
17/02/2025, 03:02 awesome-generative-ai-guide/interview_prep/60_gen_ai_questions.md at main · aishwaryanr/awesome-generative-ai-guide · GitHub

Answer
Embeddings are dense, low-dimensional representations of high-dimensional
data, serving as a fundamental concept in machine learning to efficiently capture
the essence of data entities (such as words, sentences, or images) in a form that
computational models can process. Unlike raw input data, which might be sparse
and high-dimensional (e.g., one-hot encoded vectors for words), embeddings
map these entities to continuous vectors, preserving semantic relationships while
significantly reducing dimensionality. This compact representation enables
models to perform operations and learn patterns more effectively, capturing
similarities and differences in the underlying data. For instance, in natural
language processing, word embeddings place semantically similar words closer
in the embedding space, facilitating a more nuanced understanding of language
by machine learning models.

2. Compare and contrast word embeddings and sentence embeddings. How do

their applications differ, and what considerations come into play when choosing
between them?
Answer:
Word Embeddings:
Scope: Represent individual words as vectors, capturing semantic meanings
based on usage context.
Applications: Suited for word-level tasks like synonym detection, part-of-
speech tagging, and named entity recognition.
Characteristics: Offer static representations where each word has one
embedding, potentially limiting their effectiveness for words with multiple
meanings.
Sentence Embeddings:
Scope: Extend the embedding concept to entire sentences or longer texts,
aiming to encapsulate the overall semantic content.
Applications: Used for tasks requiring comprehension of broader contexts,
such as document classification, semantic text similarity, and sentiment
analysis.
Characteristics: Provide dynamic representations that consider word
interactions and sentence structure, better capturing the context and
nuances of language use.
Considerations for Choosing Between Them:
[Link] 37/47
17/02/2025, 03:02 awesome-generative-ai-guide/interview_prep/60_gen_ai_questions.md at main · aishwaryanr/awesome-generative-ai-guide · GitHub

Task Requirements: Word embeddings are preferred for analyzing linguistic

features at the word level, while sentence embeddings are better for tasks
involving understanding of sentences or larger text units.
Contextual Sensitivity: Sentence embeddings or contextual word
embeddings (like BERT) are more adept at handling the varying meanings of
words across different contexts.
Computational Resources: Generating and processing sentence
embeddings, especially from models like BERT, can be more resource-
intensive.
Data Availability: The effectiveness of embeddings correlates with the
diversity and size of the training data.
The decision between word and sentence embeddings hinges on the specific
needs of the NLP task, the importance of context, computational considerations,
and the nature of the training data. Each type of embedding plays a crucial role in
NLP, and their effective use is key to solving various linguistic challenges.

3. Explain the concept of contextual embeddings. How do models like BERT

generate contextual embeddings, and in what scenarios are they advantageous
compared to traditional word embeddings?
Answer:
Contextual embeddings are dynamic representations of words that change based
on the word's context within a sentence, offering a more nuanced understanding
of language. Models like BERT generate contextual embeddings by using a deep
transformer architecture, processing the entire sentence at once, allowing the
model to capture the relationships and dependencies between words.
Advantages: Contextual embeddings excel over traditional, static word
embeddings in tasks requiring a deep understanding of context, such as
sentiment analysis, where the meaning of a word can shift dramatically based on
surrounding words, or in language ambiguity resolution tasks like homonym and
polysemy disambiguation. They provide a richer semantic representation by
considering the word's role and relations within a sentence.

4. Discuss the challenges and strategies involved in generating cross-modal

embeddings, where information from multiple modalities, such as text and image,
is represented in a shared embedding space.
Answer:
[Link] 38/47
17/02/2025, 03:02 awesome-generative-ai-guide/interview_prep/60_gen_ai_questions.md at main · aishwaryanr/awesome-generative-ai-guide · GitHub

Generating cross-modal embeddings faces several challenges, including aligning

semantic concepts across modalities with inherently different data
characteristics and ensuring the embeddings capture the essence of both
modalities. Strategies to address these challenges include:
Joint Learning: Training models on tasks that require understanding both
modalities simultaneously, encouraging the model to find a common semantic
ground.
Canonical Correlation Analysis (CCA): A statistical method to align the
embeddings from different modalities in a shared space by maximizing their
correlation.
Contrastive Learning: A technique that brings embeddings of similar items
closer together while pushing dissimilar items apart, applied across modalities to
ensure semantic alignment.

5. When training word embeddings, how can models be designed to effectively

capture representations for rare words with limited occurrences in the training data?
Answer:
To capture representations for rare words, models can:
Subword Tokenization: Break down rare words into smaller units (like
morphemes or syllables) for which embeddings can be learned more robustly.
Smoothing Techniques: Use smoothing or regularization techniques to borrow
strength from similar or more frequent words.
Contextual Augmentation: Increase the representation of rare words by
artificially augmenting sentences containing them in the training data.

6. Discuss common regularization techniques used during the training of

embeddings to prevent overfitting and enhance the generalization ability of
models.
Answer:
Common regularization techniques include:
L2 Regularization: Adds a penalty on the magnitude of embedding vectors,
encouraging them to stay small and preventing overfitting to specific training
examples.
[Link] 39/47
17/02/2025, 03:02 awesome-generative-ai-guide/interview_prep/60_gen_ai_questions.md at main · aishwaryanr/awesome-generative-ai-guide · GitHub

Dropout: Randomly zeroes elements of the embedding vectors during training,

forcing the model to rely on a broader context rather than specific embeddings.
Noise Injection: Adds random noise to embeddings during training, enhancing
robustness and generalization by preventing reliance on precise values.

7. How can pre-trained embeddings be leveraged for transfer learning in downstream

tasks, and what advantages does transfer learning offer in terms of embedding
generation?
Answer:
Pre-trained embeddings, whether for words, sentences, or even larger textual
units, are a powerful resource in the machine learning toolkit, especially for tasks
in natural language processing (NLP). These embeddings are typically generated
from large corpora of text using models trained on a wide range of language
understanding tasks. When leveraged for transfer learning, pre-trained
embeddings can significantly enhance the performance of models on
downstream tasks, even with limited labeled data.
Leveraging Pre-trained Embeddings for Transfer Learning:
Initialization: In this approach, pre-trained embeddings are used to initialize
the embedding layer of a model before training on a specific downstream
task. This gives the model a head start by providing it with rich
representations of words or sentences, encapsulating a broad understanding
of language.
Feature Extraction: Here, pre-trained embeddings are used as fixed
features for downstream tasks. The embeddings serve as input to further
layers of the model that are trained to accomplish specific tasks, such as
classification or entity recognition. This approach is particularly useful when
the downstream task has relatively little training data.
Pre-trained embeddings can be directly used or fine-tuned in downstream tasks,
leveraging the general linguistic or semantic knowledge they encapsulate. This
approach offers several advantages:
Efficiency: Significantly reduces the amount of data and computational
resources needed to achieve high performance on the downstream task.
Generalization: Embeddings trained on large, diverse datasets provide a broad
understanding of language or visual concepts, enhancing the model's
generalization ability.

[Link] 40/47
17/02/2025, 03:02 awesome-generative-ai-guide/interview_prep/60_gen_ai_questions.md at main · aishwaryanr/awesome-generative-ai-guide · GitHub

Quick Adaptation: Allows models to quickly adapt to specific tasks by fine-

tuning, speeding up development cycles and enabling more flexible applications.

8. What is quantization in the context of embeddings, and how does it contribute to

reducing the memory footprint of models while preserving representation
quality?
Answer:
Quantization involves converting continuous embedding vectors into a discrete,
compact format, typically by reducing the precision of the numbers used to
represent each component of the vectors. This process significantly reduces the
memory footprint of the embeddings and the overall model by allowing the
storage and computation of embeddings in lower-precision formats without
substantially compromising their quality. Typically, embeddings are stored as 32-
bit floating-point numbers. Quantization involves converting these high-precision
embeddings into lower-precision formats, such as 16-bit floats (float16) or even
8-bit integers (int8), thereby reducing the model's memory footprint.
Quantization is particularly beneficial for deploying large-scale models on
resource-constrained environments, such as mobile devices or in browser
applications, enabling faster loading times and lower memory usage.

9. When dealing with high-cardinality categorical features in tabular data, how

would you efficiently implement and train embeddings using a neural network to
capture meaningful representations?
Answer:
For high-cardinality categorical features, embeddings can be efficiently
implemented and trained by:
Embedding Layers: Introducing embedding layers in the neural network
specifically designed to convert high-cardinality categorical features into dense,
low-dimensional embeddings.
Batch Training: Utilizing mini-batch training to efficiently handle large datasets
and high-cardinality features by processing a subset of data at a time.
Regularization: Applying regularization techniques to prevent overfitting,
especially important for categories with few occurrences.

[Link] 41/47
17/02/2025, 03:02 awesome-generative-ai-guide/interview_prep/60_gen_ai_questions.md at main · aishwaryanr/awesome-generative-ai-guide · GitHub

10. When dealing with large-scale embeddings, propose and implement an efficient
method for nearest neighbor search to quickly retrieve similar embeddings from
a massive database.
Answer
For efficient nearest neighbor search in large-scale embeddings, methods such
as approximate nearest neighbor (ANN) algorithms can be used. Techniques like
locality-sensitive hashing (LSH), tree-based partitioning (e.g., KD-trees, Ball
trees), or graph-based approaches (e.g., HNSW) enable fast retrieval by
approximating the nearest neighbors without exhaustively comparing every pair
of embeddings. Implementing these methods involves constructing an index
from the embeddings that can quickly narrow down the search space for
potential neighbors.

11. In scenarios where an LLM encounters out-of-vocabulary words during

embedding generation, propose strategies for handling such cases.
Answer:
To handle out-of-vocabulary (OOV) words, strategies include:
Subword Tokenization: Breaking down OOV words into known subwords or
characters and aggregating their embeddings.
Zero or Random Initialization: Assigning a zero or randomly generated vector for
OOV words, optionally fine-tuning these embeddings if training data is available.
Fallback to Similar Words: Using embeddings of semantically or morphologically
similar words as a proxy for OOV words.

12. Propose metrics for quantitatively evaluating the quality of embeddings

generated by an LLM. How can the effectiveness of embeddings be assessed in
tasks like semantic similarity or information retrieval?
Answer:
Quality of embeddings can be evaluated using metrics such as:
Cosine Similarity: Measures the cosine of the angle between two embedding
vectors, useful for assessing semantic similarity.
Precision@k and Recall@k for Information Retrieval: Evaluates how many of the
top-k retrieved documents (or embeddings) are relevant to a query.
[Link] 42/47
17/02/2025, 03:02 awesome-generative-ai-guide/interview_prep/60_gen_ai_questions.md at main · aishwaryanr/awesome-generative-ai-guide · GitHub

Word Embedding Association Test (WEAT): Assesses biases in embeddings by

measuring associations between sets of target words and attribute words.

13. Explain the concept of triplet loss in the context of embedding learning.
Answer
Triplet loss is used to learn embeddings by ensuring that an anchor embedding is
closer to a positive embedding (similar content) than to a negative embedding
(dissimilar content) by a margin. This loss function helps in organizing the
embedding space such that embeddings of similar instances cluster together,
while embeddings of dissimilar instances are pushed apart, enhancing the
model's ability to discriminate between different categories or concepts.

14. In loss functions like triplet loss or contrastive loss, what is the significance of
the margin parameter?
Answer:
The margin parameter in triplet or contrastive loss functions specifies the desired
minimum difference between the distances of positive and negative pairs to the
anchor. Adjusting the margin impacts the strictness of the separation enforced in
the embedding space, influencing both the learning process and the quality of
the resulting embeddings. A larger margin encourages embeddings to be spread
further apart, potentially improving the model's discrimination capabilities, but if
set too high, it might lead to training difficulties or degraded performance due to
an overly stringent separation criterion.

Training, Inference and Evaluation

1. Discuss challenges related to overfitting in LLMs during training. What strategies
and regularization techniques are effective in preventing overfitting, especially
when dealing with massive language corpora?
Answer:
Challenges: Overfitting in Large Language Models can lead to models that
perform well on training data but poorly on unseen data. This is particularly
challenging with massive language corpora, where the model may memorize
rather than generalize.
Strategies and Techniques:
[Link] 43/47
17/02/2025, 03:02 awesome-generative-ai-guide/interview_prep/60_gen_ai_questions.md at main · aishwaryanr/awesome-generative-ai-guide · GitHub

Data Augmentation: Increases the diversity of the training set, helping models to
generalize better.
Regularization: Techniques such as dropout, L2 regularization, and early
stopping can discourage the model from memorizing the training data.
Model Simplification: Although challenging for LLMs, reducing model complexity
can mitigate overfitting.
Batch Normalization: Helps in stabilizing the learning process and can
contribute to preventing overfitting.

2. Large Language Models often require careful tuning of learning rates. How do you
adapt learning rates during training to ensure stable convergence and efficient
learning for LLMs?
Answer:
Adapting Learning Rates:
Learning Rate Scheduling: Gradually reducing the learning rate during training
can help in achieving stable convergence. Techniques like step decay,
exponential decay, or cosine annealing are commonly used.
Adaptive Learning Rate Algorithms: Methods such as Adam or RMSprop
automatically adjust the learning rate based on the training process, improving
efficiency and stability.

3. When generating sequences with LLMs, how can you handle long context
lengths efficiently? Discuss techniques for managing long inputs during real-time
inference.
Answer:
Some solutions are:
Fine-tuning on Longer Contexts: Training a model on shorter sequences and
then fine-tuning it on longer sequences may seem like a solution. However,
this approach may not work well with the original Transformer due to
Positional Sinusoidal Encoding limitations.
Flash Attention: FlashAttention optimizes the attention mechanism for GPUs
by breaking computations into smaller blocks, reducing memory transfer
overheads and enhancing processing speed.
[Link] 44/47
17/02/2025, 03:02 awesome-generative-ai-guide/interview_prep/60_gen_ai_questions.md at main · aishwaryanr/awesome-generative-ai-guide · GitHub

Multi-Query Attention (MQA): MQA is an optimization over the standard

Multi-Head Attention (MHA), sharing a common weight matrix for projecting
"key" and "value" across heads, leading to memory efficiency and faster
inference speed.
Positional Interpolation (PI): Adjusts position indices to fit within the existing
context size using mathematical interpolation techniques.
Rotary Positional Encoding (RoPE): Rotates existing embeddings based on
their positions, capturing sequence position in a more fluid manner.
ALiBi (Attention with Linear Biases): Enhances the Transformer's adaptability
to varied sequence lengths by introducing biases in the attention
mechanism, optimizing performance on extended contexts.
Sparse Attention: Considers only some tokens within the content size when
calculating attention scores, making computation linear with respect to input
token size.

4. What evaluation metrics can be used to judge LLM generation quality

Answer:
Common metrics used to evaluate Language Model performance include:
i. Perplexity: Measures how well the model predicts a sample of text. Lower
perplexity values indicate better performance.
ii. Human Evaluation: Involves enlisting human evaluators to assess the quality
of the model's output based on criteria like relevance, fluency, coherence,
and overall quality.
iii. BLEU (Bilingual Evaluation Understudy): A metric primarily used in machine
translation tasks. It compares the generated output with reference
translations and measures their similarity.
iv. ROUGE (Recall-Oriented Understudy for Gisting Evaluation): Used for
evaluating the quality of summaries. It compares generated summaries with
reference summaries and calculates precision, recall, and F1-score.
v. Diversity: Measures the variety and uniqueness of generated responses,
often analyzed using metrics such as n-gram diversity or semantic similarity.
Higher diversity scores indicate more diverse and unique outputs.
vi. Truthfulness Evaluation: Evaluating the truthfulness of LLMs involves
techniques like comparing LLM-generated answers with human answers,
benchmarking against datasets like TruthfulQA, and training true/false
classifiers on LLM hidden layer activations.

[Link] 45/47
17/02/2025, 03:02 awesome-generative-ai-guide/interview_prep/60_gen_ai_questions.md at main · aishwaryanr/awesome-generative-ai-guide · GitHub

5. Hallucination in LLMs a known issue, how can you evaluate and mitigate it?
Answer:
Some approaches to detect and mitigate hallucinations (source):
i. Log Probability (Seq-Logprob):
Introduced in the paper "Looking for a Needle in a Haystack" by
Guerreiro et al. (2023).
Utilizes length-normalized sequence log-probability to assess the
confidence of the model's output.
Effective for evaluating translation quality and detecting hallucinations,
comparable to reference-based methods.
Offers simplicity and ease of computation during the translation
process.
ii. Sentence Similarity:
Proposed in the paper "Detecting and Mitigating Hallucinations in
Machine Translation" by David et al. (Dec 2022).
Evaluates the percentage of source contribution to generated
translations and identifies hallucinations by detecting low source
contribution.
Utilizes reference-based, internal measures, and reference-free
techniques along with measures of semantic similarity between
sentences.
Techniques like LASER, LaBSE, and XNLI significantly improve detection
and mitigation of hallucinations, outperforming previous approaches.
iii. SelfCheckGPT:
Introduced in the paper "SelfCheckGPT: Zero-Resource Black-Box
Hallucination Detection for Generative Large Language Models" by
Manakul et al. (2023).
Evaluates hallucinations using GPT when output probabilities are
unavailable, commonly seen in black-box scenarios.
Utilizes variants such as SelfCheckGPT with BERTScore and
SelfCheckGPT with Question Answering to assess informational
consistency.
Combination of different SelfCheckGPT variants provides
complementary outcomes, enhancing the detection of hallucinations.
iv. GPT4 Prompting:
Explored in the paper "Evaluating the Factual Consistency of Large
Language Models Through News Summarization" by Tam et al. (2023).
Focuses on summarization tasks and surveys different prompting
techniques and models to detect hallucinations in summaries.
[Link] 46/47
17/02/2025, 03:02 awesome-generative-ai-guide/interview_prep/60_gen_ai_questions.md at main · aishwaryanr/awesome-generative-ai-guide · GitHub

Techniques include chain-of-thought prompting and sentence-by-

sentence prompting, comparing various LLMs and baseline approaches.
Few-shot prompts and combinations of prompts improve the
performance of LLMs in detecting hallucinations.
v. G-EVAL:
Proposed in the paper "G-Eval: NLG Evaluation using GPT-4 with Better
Human Alignment" by Liu et al. (2023).
Provides a framework for LLMs to evaluate the quality of Natural
Language Generation (NLG) using chain of thoughts and form filling.
Outperforms previous approaches by a significant margin, particularly
effective for summarization and dialogue generation datasets.
Combines prompts, chain-of-thoughts, and scoring functions to assess
hallucinations and coherence in generated text.

6. What are mixture of experts models?

Answer:
Mixture of Experts (MoE) models consist of several specialized sub-models
(experts) and a gating mechanism that decides which expert to use for a given
input. This architecture allows for handling complex problems by dividing them
into simpler, manageable tasks, each addressed by an expert in that area.

7. Why might over-reliance on perplexity as a metric be problematic in evaluating

LLMs? What aspects of language understanding might it overlook?
Answer
Over-reliance on perplexity can be problematic because it primarily measures
how well a model predicts the next word in a sequence, potentially overlooking
aspects such as coherence, factual accuracy, and the ability to capture nuanced
meanings or implications. It may not fully reflect the model's performance on
tasks requiring deep understanding or creative language use.

[Link] 47/47