Here is a concise answer suitable for a 5-mark question on the challenges in generative
models:
✅ Challenges in Generative Models:
1. Mode Collapse
○ The model generates limited varieties of outputs, ignoring large parts of the
data distribution.
○ Common in GANs where the generator produces very similar outputs
regardless of input.
2. Training Instability
○ Generative models, especially GANs, can be highly sensitive to
hyperparameters and architecture choices.
○ The adversarial nature can lead to non-convergence or oscillating behavior.
3. High Computational Cost
○ Training large generative models requires significant GPU resources and
time.
○ Sampling from complex models (like autoregressive or diffusion models) can
be slow.
4. Evaluation Difficulty
○ Unlike classifiers, it is hard to objectively evaluate the quality of generated
outputs.
○ Metrics like Inception Score or FID are approximations and may not reflect
real-world usefulness.
5. Data Bias and Ethics
○ Generative models can amplify biases present in training data, producing
unfair or offensive content.
○ Managing copyright, misinformation, and deepfakes is a growing concern.
Here’s a clear comparison table differentiating Generative and Discriminative models:
Aspect Generative Models Discriminative Models
Definition Models joint probability P(X, Y) Models conditional probability
**P(Y
Purpose Learn how data is generated Learn boundaries between
classes
Goal Generate new data points Classify or predict labels
Data Usage Understand both input X and Focus mainly on output Y
output Y given X
Examples Naive Bayes, Hidden Markov Logistic Regression, SVM,
Model, GANs, VAEs Neural Networks
Can Generate ✅ Yes ❌ No
Data?
Performance Lower classification accuracy than Higher accuracy in
(usually) discriminative classification tasks
Training Often more complex Comparatively simpler
Complexity
Overfitting Risk Higher (due to modeling full Lower
distribution)
Use Cases Image generation, speech Spam detection, sentiment
synthesis analysis
Here is a clear comparison table differentiating Virtual Reality (VR) and Augmented
Reality (AR):
Aspect Virtual Reality (VR) Augmented Reality (AR)
Definition Creates a completely Overlays digital content onto the
immersive digital environment real world
Environment Fully virtual; real world is Real world remains visible and
blocked enhanced
Device Used VR headsets (e.g., Oculus Rift, AR glasses, smartphones, tablets
HTC Vive) (e.g., HoloLens, ARKit)
User Interaction Interacts within the virtual world Interacts with real world + virtual
elements
Real World Completely replaces real Enhances the real environment
Presence surroundings
Applications Gaming, training simulations, Navigation, retail, education,
virtual tours industrial maintenance
Hardware Requires powerful hardware Can work on smartphones and
Requirement and sensors lighter devices
Level of High – isolates the user from Low to medium – blends digital
Immersion the real world with reality
Examples VR racing game, virtual surgery Pokémon Go, IKEA Place app,
simulation Snapchat filters
Q) CycleGAN
🌀 Short Note on CycleGAN
CycleGAN (Cycle-Consistent Generative Adversarial Network) is a type of generative
model designed for image-to-image translation without needing paired data.
🔑 Key Concepts:
● Unpaired Translation: Unlike traditional GANs that need matching image pairs (e.g.,
horse → zebra), CycleGAN learns from unpaired datasets (a collection of horse
images and a separate collection of zebra images).
● Cycle Consistency: It ensures that translating an image to the target domain and
back again should give the original image. This is done using two generators:
○ G:X→YG: X \rightarrow Y
○ F:Y→XF: Y \rightarrow X
○ With cycle consistency loss:
F(G(x))≈xF(G(x)) \approx x and G(F(y))≈yG(F(y)) \approx y
● Discriminators: Two discriminators are used:
○ DXD_X for domain X (e.g., horses)
○ DYD_Y for domain Y (e.g., zebras)
📷 Applications:
● Style transfer (e.g., turning photos into Van Gogh-style art)
● Day ↔ Night conversion
● Photo ↔ Painting translation
● Horse ↔ Zebra transformation
CycleGAN has two generators—G and F.
● G transforms images from domain X (e.g., photos) to domain Y (e.g., artwork).
● F transforms images from domain Y back to domain X.
Discriminators:
There are two discriminators—Dₓ and Dᵧ.
● Dₓ distinguishes between real images from X and generated images from F(y).
● Dᵧ distinguishes between real images from Y and generated images from G(x)
Q) DCGAN
🔷 Short Note on DCGAN (Deep Convolutional GAN)
DCGAN stands for Deep Convolutional Generative Adversarial Network. It is a type of
GAN architecture that uses convolutional neural networks (CNNs) to generate realistic
images.
🔑 Key Features:
● Uses CNNs instead of fully connected layers to better capture spatial hierarchies in
image data.
● Introduced architectural guidelines like:
○ No pooling layers: Use strided convolutions and transposed convolutions.
○ Batch normalization in both generator and discriminator for stable training.
○ Leaky ReLU in the discriminator and ReLU in the generator (except output
layer).
● Generator learns to produce images from random noise (latent space).
● Discriminator tries to distinguish between real and fake images.
🧠 Why It’s Important:
DCGAN is foundational in generative deep learning. It improved training stability and
image quality compared to earlier GANs, and is often used as a baseline for research.
📷 Applications:
● Generating realistic faces
● Artwork creation
● Data augmentation
● Style blending
