Module 1: Fundamentals of Neural Network
What is Deep Learning?
→ Subset of ML using neural networks with many layers.
- Learns from large data.
- Uses multiple hidden layers.
- Improves automatically.
What is a neuron?
→ Computes output using weights, bias, and activation.
- Basic unit of NN.
- Processes one or more inputs.
- Produces single output.
What is MLP?
→ Multi-Layer Perceptron with multiple hidden layers.
- Fully connected.
- Used for classification and regression.
- Trained with backpropagation.
Difference between SNN and DNN?
→ SNN has few layers, DNN has many.
- DNN more complex.
- SNN used for simple tasks.
- DNN better for feature learning.
What is Gradient Descent?
→ Optimization method to reduce loss.
- Updates weights iteratively.
- Moves in direction of minimum loss.
- Used in all DL models.
Module 2: Optimization and Regularization
What is Backpropagation?
→ Algorithm to train NN by minimizing loss.
- Computes error gradient.
- Updates weights backward.
- Uses chain rule.
What are Activation Functions?
→ Introduce non-linearity into the model.
- Examples: ReLU, Sigmoid, Tanh.
- Decide neuron output.
- Helps model learn complex patterns.
What is Regularization?
→ Prevents overfitting by controlling model complexity.
- L1 and L2 methods.
- Dropout removes random neurons.
- Improves generalization.
What is Batch Normalization?
→ Normalizes inputs between layers.
- Faster convergence.
- Stabilizes training.
- Reduces internal covariate shift.
Module 3: Autoencoders
What is Autoencoder?
→ Learns to reconstruct input data.
- Has encoder and decoder.
- Used for feature learning.
- Unsupervised learning type.
What is Denoising Autoencoder?
→ Removes noise from input.
- Trained with noisy data.
- Improves data quality.
- Used in image restoration.
What is Sparse Autoencoder?
→ Few active neurons in hidden layer.
- Learns efficient representations.
- Adds sparsity constraint.
- Useful for anomaly detection.
Module 4: Convolutional Neural Networks (CNN)
What is CNN?
→ Convolutional Neural Network used for images.
- Extracts spatial features.
- Has convolution and pooling layers.
- Famous in vision tasks.
What is Padding and Stride?
→ Used to control feature map size.
- Padding adds border pixels.
- Stride defines step size.
- Affects output dimensions.
Name CNN Architectures.
→ LeNet, AlexNet, ResNet.
- LeNet – simple model.
- AlexNet – deep CNN.
- ResNet – skip connections.
Module 5: Recurrent Neural Networks (RNN)
What is RNN?
→ Processes sequential/time-series data.
- Uses memory of previous states.
- Good for text/speech.
- Handles variable-length input.
What is LSTM?
→ Long Short-Term Memory solves vanishing gradient.
- Has gates to control flow.
- Maintains long-term memory.
- Used in NLP, translation.
What is GRU?
→ Simpler version of LSTM.
- Combines forget and input gates.
- Faster to train.
- Performs similarly to LSTM.
Module 6: Generative Models and Applications
What is GAN?
→ Generative Adversarial Network generates new data.
- Has Generator and Discriminator.
- Trains in competition.
- Used for fake image creation.
What is Generator?
→ Creates synthetic data similar to real.
- Learns to fool discriminator.
- Takes random noise as input.
- Produces images or data.
Applications of GAN?
→ Used in image generation and Deepfake.
- Art creation.
- Face swapping.
- Super-resolution.