UNIT 6: Neural Network

Neural Network

MCQs

1. What is a neural network?

A) A set of interconnected resistors
B) A machine learning model inspired by the human brain
C) A network of physical neurons in the brain
D) A rule-based expert system
Answer: B

2. What are the three main layers of a neural network?
   A) Input, Processing, Storage
   B) Input, Hidden, Output
   C) Training, Testing, Validation
   D) Nodes, Edges, Weights
   Answer: B
3. What is the primary function of the input layer in a neural network?
   A) Store data permanently
   B) Process and classify data
   C) Receive and pass input data to the next layer
   D) Generate output predictions
   Answer: C
4. What is the role of weights in a neural network?
   A) Control activation functions
   B) Determine the importance of input features
   C) Define the number of layers in the network
   D) Store training data
   Answer: B
5. Which function is used to determine whether a neuron is activated?
   A) Cost function
   B) Activation function
   C) Loss function
   D) Training function
   Answer: B
6. Which of the following is NOT a common activation function?
   A) Sigmoid
   B) ReLU
   C) Tanh
   D) Linear Regression
   Answer: D
7. What is the main purpose of bias in a neural network?
   A) Improve the learning rate
   B) Offset the weighted sum to help with non-linearity
   C) Remove errors from training data
   D) Increase network complexity
   Answer: B
8. What is backpropagation?
   A) A technique to remove redundant neurons
   B) A method to adjust weights by minimizing errors
   C) A way to generate random predictions
   D) A visualization tool for deep learning models
   Answer: B
9. Which neural network is best suited for image recognition?
   A) Feedforward Neural Network
   B) Convolutional Neural Network (CNN)
   C) Recurrent Neural Network (RNN)
   D) Generative Adversarial Network (GAN)
   Answer: B
10. What is the distinguishing feature of Recurrent Neural Networks (RNNs)?
    A) Their ability to process sequential data
    B) Their efficiency in recognizing images
    C) Their use of filters to extract spatial features
    D) Their application in unsupervised learning
    Answer: A
11. What does a deep neural network (DNN) contain?
    A) One hidden layer
    B) More than three layers including input and output layers
    C) Only an input and output layer
    D) A fixed number of neurons
    Answer: B
12. Which algorithm is commonly used for optimization in neural networks?
    A) K-means clustering
    B) Gradient descent
    C) Decision trees
    D) Support vector machines
    Answer: B
13. In neural networks, what is overfitting?
    A) A network learning patterns that are too complex
    B) A network that learns only simple relationships
    C) A model that generalizes well
    D) A model with too few neurons
    Answer: A
14. What is the primary advantage of using CNNs over FNNs?
    A) CNNs can process sequential data
    B) CNNs are more efficient in recognizing spatial patterns
    C) CNNs do not require activation functions
    D) CNNs do not use weights and biases
    Answer: B
15. Which component helps a neural network “learn”?
    A) Neurons
    B) Activation functions
    C) Weights and biases
    D) Hidden layers
    Answer: C
16. What is dropout in neural networks?
    A) A method to remove outliers in training data
    B) A regularization technique to prevent overfitting
    C) A function to speed up training
    D) A way to optimize activation functions
    Answer: B
17. What is the purpose of an epoch in training neural networks?
    A) A single pass through the entire training dataset
    B) The process of tuning hyperparameters
    C) A measure of network complexity
    D) The number of neurons in a layer
    Answer: A
18. What is the primary application of Generative Adversarial Networks (GANs)?
    A) Predicting stock prices
    B) Generating synthetic data
    C) Language translation
    D) Time series forecasting
    Answer: B
19. Which tool is commonly used for building neural networks?
    A) TensorFlow
    B) Excel
    C) Tableau
    D) SQL
    Answer: A
20. What is a perceptron?
    A) A basic unit of deep learning
    B) A single-layer neural network
    C) A recurrent neural network
    D) A convolutional layer
    Answer: B
21. What does the loss function in a neural network measure?
    A) The network’s accuracy
    B) The error between predicted and actual values
    C) The number of training examples
    D) The total number of neurons
    Answer: B
22. Which type of neural network is best for natural language processing?
    A) CNN
    B) FNN
    C) RNN
    D) Perceptron
    Answer: C
23. What is the key feature of a feedforward neural network?
    A) It has cyclic connections
    B) Data flows only in one direction
    C) It uses backpropagation
    D) It operates without activation functions
    Answer: B
24. What type of dataset is best suited for RNNs?
    A) Images
    B) Time-series data
    C) Tabular data
    D) Static graphs
    Answer: B
25. Which of the following best describes the term “training data”?
    A) The dataset used for final evaluation
    B) The data used to adjust weights and biases
    C) A dataset stored in a neural network
    D) The output of a trained model
    Answer: B

Important Questions with answer:

1. What is a neural network, and how does it work?

Answer: A neural network is a machine learning model inspired by the structure of the human brain. It consists of interconnected nodes (neurons) organized in layers: an input layer, one or more hidden layers, and an output layer. Each neuron receives input, processes it using weights and an activation function, and passes the result to the next layer. The network is trained using algorithms like backpropagation, adjusting weights to minimize error and improve predictions over time.

---

2. Explain the importance of activation functions in neural networks.

Answer: Activation functions introduce non-linearity to a neural network, allowing it to model complex patterns in data. Without them, the network would behave like a linear regression model, unable to capture intricate relationships. Common activation functions include:

• Sigmoid: Maps inputs between 0 and 1, useful for binary classification.
• ReLU (Rectified Linear Unit): Helps avoid vanishing gradient issues by setting negative values to zero.
• Tanh: Similar to sigmoid but outputs between -1 and 1, allowing stronger gradients for learning.

---

3. What are the key components of a neural network?

Answer: A neural network consists of:

1. Neurons (Nodes): The fundamental units that process and transmit information.
2. Weights: Values assigned to connections between neurons to determine importance.
3. Bias: A constant added to a neuron’s weighted sum before activation.
4. Layers:
   • Input layer: Receives raw data.
   • Hidden layers: Process data through transformations.
   • Output layer: Produces final predictions.
5. Activation Functions: Decide whether a neuron should be activated based on inputs.
6. Loss Function: Measures the network’s prediction error.
7. Optimizer: Adjusts weights to reduce loss.

---

4. Explain the difference between a feedforward neural network and a recurrent neural network.

Answer:

• Feedforward Neural Network (FNN): Data flows in one direction, from input to output. There are no loops or feedback connections, making it suitable for static data like image classification.
• Recurrent Neural Network (RNN): Designed for sequential data, RNNs use feedback loops to retain information from previous steps, making them ideal for time-series analysis, speech recognition, and language modeling.

---

5. How does backpropagation work in training a neural network?

Answer: Backpropagation is an optimization algorithm used to train neural networks. It consists of:

1. Forward Propagation: Inputs pass through the network, and an output is generated.
2. Loss Calculation: The difference between the predicted and actual values is measured using a loss function.
3. Backward Propagation: The error is propagated back through the layers to adjust weights using gradient descent.
4. Weight Update: The optimizer updates the weights to reduce the loss. This process repeats until the network achieves satisfactory accuracy.

---

6. What are convolutional neural networks (CNNs), and why are they used in image processing?

Answer: CNNs are a type of neural network designed for image recognition tasks. They use convolutional layers to extract spatial features, pooling layers to reduce dimensionality, and fully connected layers for classification. CNNs efficiently detect patterns such as edges, shapes, and textures, making them ideal for applications like facial recognition, object detection, and medical imaging.

---

7. What are generative adversarial networks (GANs), and how do they work?

Answer: GANs are a type of neural network used for generating synthetic data. They consist of two networks:

• Generator: Creates new data samples.
• Discriminator: Evaluates if the generated data is real or fake.
  Both networks train simultaneously, with the generator improving its output to fool the discriminator. GANs are used for image generation, style transfer, and data augmentation.

---

8. What are some common challenges in training deep neural networks?

Answer:

1. Overfitting: The network memorizes training data instead of generalizing.
2. Vanishing/Exploding Gradients: Gradient values become too small (vanish) or too large (explode), hindering learning.
3. High Computational Cost: Training deep networks requires substantial processing power and memory.
4. Hyperparameter Tuning: Selecting optimal parameters like learning rate and batch size can be complex.
5. Lack of Training Data: Deep networks require large datasets for accurate predictions.

---

9. What is the significance of the learning rate in neural networks?

Answer: The learning rate controls how much the weights are updated during training.

• Too high: The network may not converge and overshoot the optimal solution.
• Too low: Training becomes slow, and the network may get stuck in local minima.
  Optimizers like Adam and SGD help adjust the learning rate dynamically.

---

10. How do neural networks handle non-linearity?

Answer: Neural networks introduce non-linearity through activation functions like ReLU, Sigmoid, and Tanh. Without these functions, the network would only be capable of learning linear relationships, limiting its ability to capture complex patterns in data.

---

11. What is dropout, and why is it used in neural networks?

Answer: Dropout is a regularization technique where random neurons are ignored during training. This prevents overfitting by ensuring that the network does not become too reliant on specific neurons, improving generalization on unseen data.

---

12. How do Recurrent Neural Networks (RNNs) process sequential data?

Answer: RNNs retain information from previous inputs using a feedback loop. They maintain a hidden state that updates at each time step, allowing them to recognize patterns over time. This makes RNNs useful for tasks like speech recognition and time-series forecasting.

---

13. How does batch normalization improve neural network training?

Answer: Batch normalization normalizes the input of each layer, reducing internal covariate shift. This stabilizes training, speeds up convergence, and allows for higher learning rates, improving model accuracy.

---

14. What is transfer learning, and why is it useful?

Answer: Transfer learning involves using a pre-trained model on a new task. Instead of training from scratch, an existing model trained on a large dataset (e.g., ImageNet) is fine-tuned for a specific application, reducing training time and improving performance with limited data.

---

15. What is the role of pooling layers in CNNs?Answer: Pooling layers reduce the spatial dimensions of feature maps, decreasing computational complexity and making the network more robust to variations in input images. Common types include max pooling and average pooling.

Scenario-Based Question on Neural Networks Application

Question 1: You are working as a data scientist for a healthcare company that wants to build an AI model to detect pneumonia from chest X-rays. Which type of neural network would you choose, and why?

Answer:
A Convolutional Neural Network (CNN) would be the best choice because it is specifically designed for image processing tasks. CNNs use convolutional layers to extract important spatial features such as edges, textures, and patterns from medical images. Pooling layers help reduce dimensionality, making computation more efficient. This enables the model to accurately detect pneumonia in chest X-rays by learning patterns in medical imaging datasets.

---

Question 2: You are building a predictive text system for a smartphone keyboard. Which type of neural network would you use and how would it work?

Answer:
A Recurrent Neural Network (RNN) would be the best choice because it is designed for sequential data processing. RNNs retain memory of previous words in a sequence through feedback loops, allowing them to predict the next word based on context. An advanced variant, Long Short-Term Memory (LSTM) networks, can handle long-term dependencies better, improving accuracy in text prediction. The model would be trained on large datasets of text to learn word relationships and sentence structures.

---

Question 3: A company has developed an AI hiring system that uses a neural network to screen job applications. However, it has been found that the system favors certain demographics over others. What could be causing this issue, and how can it be addressed?

Answer:
The issue is likely due to algorithmic bias, which occurs when the training data used for the neural network is imbalanced or reflects human biases. If the historical hiring data includes biases against certain demographics, the neural network will learn and replicate those biases.

To address this:

1. Use Diverse and Representative Data: Ensure training data includes candidates from all demographics.
2. Regular Audits: Continuously test the model for biased outcomes and retrain as needed.
3. Fairness Constraints: Implement fairness-aware learning techniques to reduce bias.
4. Transparency: Make hiring criteria explainable and interpretable to avoid hidden biases.

---

Question 4: A company has developed a neural network model that performs exceptionally well on training data but fails on real-world test data. What problem is the model facing, and how can it be resolved?

Answer:
The model is suffering from overfitting, meaning it has memorized the training data instead of learning general patterns. This reduces its ability to perform well on unseen data.

To prevent overfitting:

1. Use Dropout: Randomly deactivate neurons during training to prevent dependency on specific features.
2. Increase Training Data: More diverse data helps the model generalize better.
3. Regularization (L1/L2): Penalizes large weight values, preventing the model from becoming too complex.
4. Cross-Validation: Splitting data into multiple sets ensures the model performs consistently.

---

Question 5: You are tasked with building a temperature conversion model that predicts Fahrenheit values based on Celsius input. How would you implement this using a neural network?

Answer:
To build this model using a neural network in TensorFlow/Keras, follow these steps:

1. Prepare Training Data:
   • Create a dataset of Celsius values (X) and corresponding Fahrenheit values (Y).
   • Example: X = [-40, -10, 0, 8, 15, 22, 38], Y = [-40, 14, 32, 46, 59, 72, 100].
2. Define the Neural Network:
   • Use a simple feedforward neural network with one input layer, one hidden layer, and one output layer.
   • model = tf.keras.Sequential([tf.keras.layers.Dense(units=1, input_shape=[1])])
3. Compile the Model:
   • Use Mean Squared Error as the loss function and Adam optimizer for training.
   • model.compile(loss=’mean_squared_error’, optimizer=tf.keras.optimizers.Adam(0.1))
4. Train the Model:
   • model.fit(X, Y, epochs=500, verbose=False)
5. Make Predictions:
   • Convert new values: print(model.predict([100])) should approximate 212°F.