Deep Learning Lab-BAI701-Viva questions

 

1. What is a Neural Network?

A Neural Network is a type of artificial intelligence model inspired by how the human brain works. Just like your brain has neurons that pass signals to each other, a neural network has artificial neurons arranged in layers.
These neurons learn patterns from data.

Structure of a Neural Network

A basic neural network has three types of layers:

1. Input Layer
• Takes the data (like images, text, numbers).
2. Hidden Layers
• These layers do the actual learning.
• More hidden layers → deeper network → better pattern learning.
3. Output Layer
• Gives the final result
  (e.g., “DR present”, “cat”, “spam email”, etc.).

How it learns

It learns by adjusting the weight of each connection, similar to how we strengthen or weaken ideas in our brain.

1. Data is given to the network.
2. The network predicts something.
3. If prediction is wrong, it adjusts itself.
4. Over time, it gets better and better.

Example Uses

• Detecting diabetic retinopathy
• Face recognition
• Language translation
• Self-driving cars

2. What is Reinforcement Learning?

Reinforcement Learning (RL) is a type of machine learning where an agent learns by interacting with an environment and receiving rewards or punishments.

It is like training a pet or learning a game.

Key Idea

The agent tries different actions.

• If the action is good → It gets a reward
• If the action is bad → It gets a penalty

Over time, the agent learns the best actions to maximize reward.

Important Terms

• Agent: The learner (like a robot or software program).
• Environment: Where it acts (a game, a room, a factory).
• Action: What the agent does.
• Reward: Feedback given for actions.
• Policy: The strategy the agent learns to follow.

Example

A robot wants to learn to walk:

• When it moves forward → +10 reward
• When it falls → –5 penalty
• Over time, it learns stable walking.

Where RL is used

• Robotics (navigation, manipulation)
• Game playing (Chess, Go, Atari, etc.)
• Self-driving cars
• Healthcare treatment recommendations
• Industrial automation

 

3.Explain Autoencoder in detail

An Autoencoder is a special type of neural network that learns to compress data and then reconstruct it.
Its main goal is to learn the most important patterns in the data.

In simple words:

Autoencoder = Data Compressor + Data Rebuilder

It is an unsupervised learning technique because it does not need labeled data.

 Why do we need Autoencoders?

Autoencoders help in:

• Reducing the size of data (compression)
• Removing noise from images
• Finding important features automatically
• Detecting anomalies (fraud, defects, rare events)
• Enhancing medical images

They are widely used in image processing, NLP, and anomaly detection.

 

Structure of an Autoencoder

An autoencoder has three main parts:

1. Encoder

• Takes the input data (image, text, etc.)
• Compresses it into a smaller-sized representation
• This condensed representation is called latent vector or bottleneck

Example:

Input image: 1000 pixels
Encoder compresses it to: 50 values

It learns only the essential features.

2. Bottleneck (Latent Space)

This is the compressed knowledge of the data.

• Represents only the important features.
• Removes unnecessary details.
• Acts like the “core meaning” of the data.

This is where the autoencoder learns hidden patterns.

3. Decoder

• Takes the latent vector
• Tries to reconstruct the original input

The goal is to make the reconstructed output as close as possible to the original input.

How Autoencoders Learn

During training, the autoencoder compares:

• Input Image
  vs
• Reconstructed Image

The difference between them is the reconstruction error.

The network adjusts weights to minimize this error.

4.     Explain Word Embedding in detail

Word Embedding is a technique in Natural Language Processing (NLP) where words are converted into numerical vectors so that a machine can understand and process them.

In simple words:

Word embedding = representing words as numbers in such a way that words with similar meanings have similar vector values.

It helps computers understand meaning, context, and relationships between words.

 Why Do We Need Word Embeddings?

Computers understand numbers, not text.

Earlier, NLP used One-Hot Encoding:

• Each word gets a unique vector (0s and 1s)
• Very large and sparse
• No meaning — “king” and “queen” look unrelated

Example:

• KING = [0,0,1,0,0,0,...]
• QUEEN = [0,0,0,1,0,0,...]

No similarity is captured.

·         Problem: One-hot encoding does not capture meaning or relationships.

·         Word Embeddings solve this by placing similar words close together in vector space.

 What Are Word Embeddings?

Word embeddings are dense vectors (e.g., 50, 100, 300 dimensions) that capture:

• Semantics (meaning)
• Context
• Relationships
• Similarity between words

Example of embeddings:

• king → [0.27, 0.68, 0.12, …]
• queen → [0.26, 0.70, 0.10, …]
• man → [0.11, 0.01, 0.19, …]
• woman → [0.10, 0.02, 0.20, …]

Notice: king and queen vectors are similar, as are man and woman.

How Word Embeddings Work

Word embeddings rely on a simple idea:

“Tell me who your neighbors are, and I’ll tell you what you mean.”

Words occurring in similar contexts have similar meanings.

Example:

• “The doctor treated the patient.”
• “The nurse cared for the patient.”

The words doctor and nurse appear near patient, so embeddings place them close together.

 

5.   Explain Types of Word Embeddings

1️⃣ Word2Vec (Most popular)

Created by Google.

Two versions:

• CBOW (Continuous Bag of Words)
  Predicts a word from surrounding context
  Example: ___ ate an apple → predicts “He”
• Skip-Gram
  Predicts context words from a given word
  Example: “apple” → predicts “ate”, “fruit”, “red”

Key feature:

• Learns semantic relationships like:

👉 king – man + woman = queen

---

2️⃣ GloVe (Global Vectors)

Created by Stanford.

Uses word-word co-occurrence statistics from the entire text.

Strength:

Captures global relationships, not just local context.

---

3️⃣ FastText (by Facebook)

FastText breaks words into sub-word units (character n-grams).

Example:

• “playing” → “play”, “lay”, “ing”

Strength:

• Handles rare words better
• Works well for languages with rich morphology (Tamil, Kannada, Hindi)

---

4️⃣ Contextual Embeddings (Modern NLP)

These embeddings depend on the sentence context.

Examples:

• BERT
• GPT
• ELMo

Example:

Word: "bank"

• “He sat near the bank of the river.”
• “He deposited money in the bank.”

Traditional embeddings give same vector → WRONG
Contextual embeddings give different vectors → CORRECT

 

6.   Datasets Used in the Deep Learning Lab – Detailed Explanation

Your lab manual uses five major datasets across different deep learning tasks:

1. Custom Text Corpus → for Word Embeddings
2. MNIST Dataset → for Deep Neural Networks & Autoencoders
3. CIFAR-10 Dataset → for Pretrained CNN Models
4. IMDB Movie Reviews Dataset → for Text Classification
5. Synthetic Time Series Dataset → for LSTM Forecasting

Below is a clear explanation of each.

---

1️⃣ Custom Text Corpus (Used in Word Embedding Lab)

Where Used:

✔ Experiment 1 – Generate Word Embeddings using Word2Vec

What it is:

A text dataset created by the user, often raw text paragraphs like:

• Articles
• Documents
• Stories
• Wikipedia text
• Classroom notes

This text is fed into spaCy for tokenization and Word2Vec for generating embeddings.

Why used:

• To help students understand how Word2Vec learns context.
• Shows how embeddings capture similarity and relations between words.

Data Format:

A plain string or text file, converted into:

• Tokens → words
• Training pairs → (center word, context word)

Example tokens:
["neural", "networks", "are", "powerful", "models", ...]

The dataset size can be small, even 5–10 sentences, because goal is learning concept.

---

2️⃣ MNIST Dataset (Used in Experiments 2 & 4)

📌 Where Used:

✔ Experiment 2 – Deep Neural Network for Classification
✔ Experiment 4 – Autoencoder for Image Compression

📌 What MNIST Is:

MNIST is a world-famous dataset of handwritten digits (0–9).
It contains grayscale images, each 28×28 pixels.

Dataset size:

• 60,000 training images
• 10,000 testing images
• Total = 70,000 images

Each image:

• Resolution: 28 × 28
• Color: 1 channel (grayscale)
• Pixel range: 0–255, normalized to 0–1

Why MNIST is used:

• Very clean, standardized dataset
• Perfect for beginners
• Ideal for testing neural networks, autoencoders, and CNN basics

Where it fits in your labs:

• Deep Neural Network learns to classify digits
• Autoencoder learns to compress and reconstruct digit images

---

3️⃣ CIFAR-10 Dataset (Used in Pretrained Model Lab)

📌 Where Used:

✔ Experiment 7 – Using Pretrained MobileNetV2 for Image Classification
✔ Prediction and Evaluation on New Images

📌 What CIFAR-10 Is:

A popular dataset of color images, used to test computer vision models.

Dataset size:

• 50,000 training images
• 10,000 test images
• Total = 60,000 images

Each image:

• Resolution: 32 × 32 pixels
• Channels: 3 (RGB)
• 10 categories:

Class	Example Images
airplane	✈️
automobile	🚗
bird	🐦
cat	🐱
deer	🦌
dog	🐶
frog	🐸
horse	🐴
ship	🚢
truck	🚚

Why CIFAR-10 is used:

• More complex than MNIST
• Suitable for testing pretrained CNNs
• Helps students understand transfer learning

Where it fits in your labs:

• You resize CIFAR images to 96×96 for MobileNetV2
• Pretrained model extracts features
• Custom classifier predicts the category

---

4️⃣ IMDB Movie Reviews Dataset (Used in Text Classification Lab)

📌 Where Used:

✔ Experiment 5 – Deep learning network for text classification (sentiment analysis)

📌 What IMDB Dataset Is:

A large text dataset of movie reviews with labeled sentiment.
Each review is either:

• ⭐ Positive
• ⭐ Negative

Dataset size:

• 25,000 training reviews
• 25,000 testing reviews
• Balanced dataset (50% positive, 50% negative)

Data format:

• Reviews are already converted into integers (word indices)
• Example:
  The movie was great → [34, 12, 5, 98]

Why IMDB is used:

• Standard benchmark for sentiment analysis
• Useful for learning embeddings + LSTM/BiLSTM networks

In your lab:

• Word indices → Embedding layer (128 dimensions)
• Bidirectional LSTM extracts meaning
• Output → probability of positive vs negative

---

5️⃣ Synthetic Time Series Dataset (Used in LSTM Forecasting)

📌 Where Used:

✔ Experiment 6 – Deep Learning Model for Time Series Forecasting

📌 What the Synthetic Dataset Is:

A generated time series made of:

• Sine wave
• Trend
• Noise

Example generated data:

y = sin(0.02x) + random_noise + trend

Dataset size:

• Usually 500 data points, but can be changed.

Why Synthetic?

• Lets students practice forecasting without needing real industrial data
• Easy to visualize
• Shows the role of sequence windows

In the lab:

• Scaled using MinMaxScaler
• Prepared in sliding windows
• LSTM predicts the next value in sequence

7.What is called as pretrained model?

A pretrained model is a deep learning model that has already been trained on a very large dataset and can be reused for your own tasks.

In simple words:

A pretrained model = a model that someone else has already trained for you on a huge dataset, so you don’t have to start from zero.

 Why Pretrained Models Are Useful

Training deep models from scratch requires:

• millions of images or text samples
• huge GPU power
• days or weeks of training

Most students, researchers, and even companies do not have these resources.

So instead, we reuse models trained by experts on massive datasets like:

• ImageNet (14 million images)
• COCO dataset
• Wikipedia text
• Common Crawl data

These models already learned generic features, and we only fine-tune them.

 Example of Pretrained Models

Pretrained CNNs for Images

• MobileNetV2
• ResNet
• VGG16
• InceptionV3
• EfficientNet

These are trained on ImageNet to recognize 1000 classes.

You can reuse them to classify:

• medical images
• animals
• vehicles
• plants
  even if these images are not part of the original dataset.

Pretrained NLP Models

• BERT
• GPT
• Word2Vec
• Fast Text

These are trained on billions of words.

They already understand:

• grammar
• sentence structure
• word meaning
• context

You only fine-tune them for tasks like:

• sentiment analysis
• summarization
• translation
• chatbots