Important Questions - Detailed Answers
1. Write a short note on GoogleNet
GoogleNet is a deep learning model developed by researchers at Google. It became popular after
winning the ImageNet Large Scale Visual Recognition Challenge (ILSVRC) in 2014.
GoogleNet uses a special module called the Inception module, which applies multiple filters (1x1,
3x3, 5x5) at the same time and combines their results. This helps in extracting different features
from the image at different scales.
It has 22 layers, which makes it deep, yet it uses fewer parameters compared to earlier models like
VGGNet or AlexNet. This makes it faster and more efficient.
GoogleNet also uses techniques like 1x1 convolutions for dimensionality reduction, global average
pooling instead of fully connected layers, and dropout to prevent overfitting.
2. Write a short note on Gradient Descent
Gradient Descent is a method used to train machine learning and deep learning models. It works by
minimizing the loss (error) of the model by adjusting the weights.
Imagine you are standing on a hill and want to reach the lowest point. You take steps in the direction
where the slope is steepest downwards. These steps are the gradients.
Gradient Descent calculates the direction and size of each step using derivatives and repeats this
process until the error is minimized.
There are different types:
- Batch Gradient Descent: Uses the whole dataset.
- Stochastic Gradient Descent (SGD): Uses one example at a time.
- Mini-batch Gradient Descent: Uses a small group of data at a time.
3. Briefly explain the architecture of CNN
CNN (Convolutional Neural Network) is designed for processing image data. It has the following
layers:
1. Input Layer: Takes the input image.
2. Convolutional Layer: Applies filters to detect features like edges, corners.
3. Activation Layer (ReLU): Adds non-linearity to help the model learn complex patterns.
4. Pooling Layer: Reduces the size of the feature map using techniques like max pooling.
5. Fully Connected Layer: Flattens and connects features to the output.
6. Output Layer: Predicts the class of the image.
CNNs are very effective for image recognition and classification tasks.
4. How is Deep Learning better than Machine Learning?
Machine Learning and Deep Learning are both used to build intelligent systems, but they differ in
many ways:
| Feature | Machine Learning | Deep Learning |
|----------------------|------------------------------------------|------------------------------------------|
| Definition | Uses algorithms to learn from data | Uses neural networks with many layers
| Feature Extraction | Manual | Automatic |
| Data Requirement | Works on small/medium data | Needs large datasets |
| Training Time | Faster | Slower |
| Performance | Good for simple problems | Excellent for complex problems |
| Example | Spam filters, recommendation systems | Self-driving cars, face recognition
Deep Learning is more powerful when handling large and complex data like images, videos, and
audio.
5. List out what are the applications of Deep RNN
Deep RNNs (Recurrent Neural Networks) are useful for tasks where data comes in sequence.
Applications include:
- Speech Recognition (like Siri or Google Assistant)
- Language Translation (Google Translate)
- Text Generation (writing essays, code, etc.)
- Music Generation
- Chatbots and virtual assistants
- Time series prediction like stock prices
They can remember previous data points, making them ideal for sequential tasks.
6. Explain Generative Adversarial Networks (GANs)
GANs are models used to generate new, realistic data. They consist of two parts:
1. Generator: Creates fake data.
2. Discriminator: Tries to identify if the data is real or fake.
They work like a game where the generator tries to fool the discriminator, and the discriminator tries
to catch it. Over time, the generator becomes very good at creating realistic data.
Applications include:
- Generating realistic images (deepfakes)
- Creating art
- Enhancing image resolution
- Game content generation
7. What is the difference between autoencoder and deep autoencoder?
Autoencoders are neural networks used to compress and then reconstruct data.
- Autoencoder: Has few hidden layers, used for simple tasks like noise removal.
- Deep Autoencoder: Has multiple hidden layers, used for complex tasks like image compression
and feature extraction.
Both are unsupervised and learn to represent input data in a lower-dimensional form and then
reconstruct it.
8. Explain any two of the following: a) Sigmoid Neurons, d) LeNet
a) Sigmoid Neurons:
These neurons use a sigmoid activation function which outputs values between 0 and 1. It is helpful
for binary classification tasks.
Formula: Sigmoid(x) = 1 / (1 + e^-x)
It helps in deciding the probability of classes.
d) LeNet:
LeNet-5 is a CNN developed by Yann LeCun for digit recognition in the 1990s. It consists of
convolutional and pooling layers followed by fully connected layers.
It was used for tasks like reading postal codes and bank checks and was one of the first CNNs that
led to modern deep learning models.
