Q.1 What is the main goal of deep learning?
To design web applications
To mimic human decision-making
To manage computer networks
To perform database queries
Explanation - Deep learning aims to create algorithms that learn hierarchical representations of data to solve complex tasks similar to human intelligence.
Correct answer is: To mimic human decision-making
Q.2 Which type of neural network is mainly used for image recognition?
RNN
CNN
GAN
DBN
Explanation - Convolutional Neural Networks (CNNs) are specifically designed for spatial data like images and excel in image recognition.
Correct answer is: CNN
Q.3 What does 'ReLU' stand for?
Rectified Linear Unit
Relative Learning Utility
Recursive Linear Update
Random Learning Unit
Explanation - ReLU is an activation function that replaces negative values with zero, improving training speed and performance.
Correct answer is: Rectified Linear Unit
Q.4 Which problem do deep neural networks commonly face when layers increase?
Overfitting
Underfitting
Vanishing gradients
Data sparsity
Explanation - As networks deepen, gradients can shrink during backpropagation, making it hard to update weights in early layers.
Correct answer is: Vanishing gradients
Q.5 Which optimizer uses momentum to accelerate gradient descent?
Adam
SGD
RMSprop
Adagrad
Explanation - Stochastic Gradient Descent (SGD) with momentum helps speed up convergence by adding a fraction of the previous gradient update.
Correct answer is: SGD
Q.6 Dropout in deep learning is used for?
Data augmentation
Reducing overfitting
Improving memory usage
Increasing learning rate
Explanation - Dropout randomly ignores neurons during training, preventing the network from becoming too dependent on specific paths.
Correct answer is: Reducing overfitting
Q.7 Which neural network is best suited for sequential data like text?
CNN
GAN
RNN
Autoencoder
Explanation - Recurrent Neural Networks (RNNs) are designed for sequential data by using hidden states that capture temporal dependencies.
Correct answer is: RNN
Q.8 In backpropagation, what is propagated backward?
Input data
Error gradients
Predicted outputs
Weight matrices
Explanation - Backpropagation calculates the gradient of the loss with respect to each weight by propagating errors backward through the network.
Correct answer is: Error gradients
Q.9 What is the main difference between supervised learning and deep learning?
Deep learning does not use data
Deep learning uses hierarchical feature learning
Supervised learning uses unstructured data
Supervised learning always outperforms deep learning
Explanation - Deep learning models automatically learn hierarchical features from raw data, unlike traditional supervised learning where features are often manually engineered.
Correct answer is: Deep learning uses hierarchical feature learning
Q.10 What is a common activation function besides ReLU?
Sigmoid
Linear Regression
Softmax Regression
K-Means
Explanation - Sigmoid maps inputs between 0 and 1, making it useful for probabilities, though prone to vanishing gradients.
Correct answer is: Sigmoid
Q.11 What does the 'Softmax' function output?
Continuous values
Binary values
Class probabilities
Error rates
Explanation - Softmax outputs a probability distribution over multiple classes, commonly used in classification problems.
Correct answer is: Class probabilities
Q.12 Which deep learning framework was developed by Google?
PyTorch
Keras
TensorFlow
Caffe
Explanation - TensorFlow is Google’s open-source deep learning framework widely used for research and production.
Correct answer is: TensorFlow
Q.13 Which layer is used to reduce dimensions in CNNs?
Pooling layer
Dropout layer
Batch normalization
Fully connected layer
Explanation - Pooling layers reduce the spatial dimensions, lowering computation while retaining key features.
Correct answer is: Pooling layer
Q.14 Batch Normalization helps in?
Preventing vanishing gradients
Reducing dataset size
Increasing model size
Eliminating dropout
Explanation - Batch normalization normalizes activations, stabilizing learning and reducing vanishing gradients.
Correct answer is: Preventing vanishing gradients
Q.15 Which neural network is commonly used in speech recognition?
CNN
RNN
DBN
GAN
Explanation - RNNs capture temporal dependencies, making them suitable for sequential tasks like speech recognition.
Correct answer is: RNN
Q.16 GANs are mainly used for?
Classification
Clustering
Data generation
Regression
Explanation - Generative Adversarial Networks generate realistic synthetic data through adversarial training of generator and discriminator.
Correct answer is: Data generation
Q.17 What is the purpose of a loss function?
To optimize the learning rate
To calculate error between predictions and actual values
To select features
To normalize data
Explanation - Loss functions measure the error to guide model updates during training.
Correct answer is: To calculate error between predictions and actual values
Q.18 What is transfer learning?
Training a model from scratch
Reusing a pre-trained model on a new task
Randomly initializing weights
Converting supervised learning into unsupervised
Explanation - Transfer learning leverages knowledge from one task by reusing pre-trained models on similar tasks, saving time and resources.
Correct answer is: Reusing a pre-trained model on a new task
Q.19 Which function is typically used for binary classification?
Softmax
Sigmoid
ReLU
Tanh
Explanation - Sigmoid squashes outputs between 0 and 1, ideal for binary probability estimation.
Correct answer is: Sigmoid
Q.20 What does an autoencoder do?
Encodes labels
Learns compressed representations
Generates random noise
Increases dataset size
Explanation - Autoencoders learn efficient, lower-dimensional encodings of data through unsupervised training.
Correct answer is: Learns compressed representations
Q.21 Which of these is an unsupervised deep learning algorithm?
Autoencoder
CNN
RNN
LSTM
Explanation - Autoencoders are unsupervised because they don’t require labeled data, learning to reconstruct inputs.
Correct answer is: Autoencoder
Q.22 What does LSTM stand for?
Long Short-Term Memory
Linear Sequential Training Model
Latent Semantic Transfer Model
Large Scale Training Method
Explanation - LSTM is a type of RNN designed to overcome vanishing gradients by using gates to control information flow.
Correct answer is: Long Short-Term Memory
Q.23 Why are GPUs preferred in deep learning?
They store larger datasets
They have more RAM
They perform parallel computations efficiently
They reduce model size
Explanation - GPUs handle large matrix computations in parallel, significantly speeding up deep learning training.
Correct answer is: They perform parallel computations efficiently
Q.24 Which technique helps prevent overfitting?
Dropout
Batch size increase
Learning rate decay
All of the above
Explanation - Dropout, adjusting batch size, and learning rate scheduling all help improve generalization and reduce overfitting.
Correct answer is: All of the above
Q.25 Which component adjusts model parameters during training?
Activation function
Optimizer
Loss function
Batch size
Explanation - Optimizers update model parameters like weights using gradients from the loss function.
Correct answer is: Optimizer