Machine Learning Deep Learning and Neural Network Syllabus

What is the difference between L1 and L2 regularization?

L1 ($\lambda \sum |w_{i}|$) promotes sparsity, driving some weights exactly to zero (feature selection); L2 ($\lambda \sum w_{i}^{2}$) shrinks weights smoothly toward zero without making them exactly zero.

What is dropout and why does it reduce overfitting?

Dropout randomly deactivates a fraction $p$ of neurons during each training step, preventing co-adaptation of features and acting like training an ensemble of subnetworks, which improves generalization.

What is batch normalization and what are its benefits?

It normalizes layer inputs per mini-batch to zero mean and unit variance, then scales/shifts with learnable parameters $\gamma$ and $\beta$. Benefits: faster, more stable training, higher learning rates, and mild regularization.

What is early stopping?

A regularization technique that halts training when validation performance stops improving (begins to degrade), preventing the model from overfitting the training data.

What is the purpose of a convolutional layer in a CNN?

It applies learnable filters (kernels) that slide across the input to detect local spatial features (edges, textures, shapes) while sharing weights, drastically reducing parameters and exploiting spatial locality.

Give the output size formula for a convolution layer.

$$O = \left\lfloor \frac{W - F + 2P}{S} \right\rfloor + 1$$ where $W$ is input size, $F$ filter size, $P$ padding, and $S$ stride.

What is the role of a pooling layer (e.g. max pooling) in a CNN?

Pooling downsamples feature maps to reduce spatial dimensions and computation, provides translation invariance, and controls overfitting. Max pooling outputs the maximum value within each window.

What are the two key properties that make CNNs efficient for images?

Parameter sharing (the same filter is reused across all spatial positions) and local connectivity / sparse interactions (each neuron connects only to a local receptive field).

What is a recurrent neural network (RNN) designed for?

RNNs process sequential data by maintaining a hidden state that carries information across time steps, making them suited for time series, text, and speech where order matters.

Write the basic RNN hidden state update equation.

$$h_{t} = \phi(W_{hh} h_{t-1} + W_{xh} x_{t} + b)$$ where $h_{t}$ is the hidden state at time $t$, $x_{t}$ the input, and $\phi$ a non-linear activation.

Why do vanilla RNNs struggle with long-term dependencies?

Repeated multiplication during backpropagation through time causes gradients to vanish or explode, so the network cannot retain information across long sequences.

What problem do LSTMs solve and what are their three gates?

LSTMs solve the long-term dependency / vanishing gradient problem using a cell state and three gates: the forget gate, input gate, and output gate, which regulate information flow over time.

How does a GRU differ from an LSTM?

A GRU (Gated Recurrent Unit) uses two gates (reset and update) and merges the cell and hidden state, making it simpler and faster with fewer parameters, often achieving comparable performance to LSTMs.

What is a Generative Adversarial Network (GAN) composed of?

A GAN consists of two networks trained adversarially: a Generator $G$ that creates fake samples from noise, and a Discriminator $D$ that tries to distinguish real from fake samples.

Write the GAN minimax objective.

$$\min_{G}\max_{D} \; \mathbb{E}_{x \sim p_{\text{data}}}[\log D(x)] + \mathbb{E}_{z \sim p_{z}}[\log(1 - D(G(z)))]$$

What is mode collapse in GANs?

A failure mode where the generator produces only a limited variety of outputs (or a single output) that fool the discriminator, failing to capture the full diversity of the real data distribution.

What is transfer learning?

A technique that reuses a model pre-trained on a large source task as a starting point for a related target task, transferring learned features to reduce data and training requirements.

What is fine-tuning in the context of transfer learning?

Taking a pre-trained network and continuing training (all or some layers) on the new target dataset, usually with a small learning rate, to adapt the learned features to the new task.

When freezing layers in transfer learning, which layers are typically frozen and why?

Early/lower layers are frozen because they capture generic features (edges, textures) transferable across tasks, while later layers are retrained to learn task-specific high-level features.

What is reinforcement learning?

A learning paradigm where an agent learns to take actions in an environment to maximize cumulative reward through trial-and-error interaction, guided by a reward signal rather than labeled data.

Define the return (discounted cumulative reward) in reinforcement learning.

$$G_{t} = \sum_{k=0}^{\infty} \gamma^{k} r_{t+k+1}$$ where $\gamma \in [0,1]$ is the discount factor weighting future rewards.

Write the Q-learning update rule.

$$Q(s,a) \leftarrow Q(s,a) + \alpha\left[r + \gamma \max_{a'} Q(s',a') - Q(s,a)\right]$$ where $\alpha$ is the learning rate and $\gamma$ the discount factor.

What is the exploration-exploitation tradeoff?

The dilemma between exploiting known high-reward actions to maximize immediate reward versus exploring new actions to discover potentially better long-term strategies; often balanced with an $\epsilon$-greedy policy.

What is a word embedding in NLP?

A dense, low-dimensional vector representation of a word that captures semantic and syntactic relationships, so that similar words have nearby vectors (e.g. Word2Vec, GloVe).

What is the attention mechanism's core idea?

It lets a model dynamically weight the relevance of different input elements when producing each output, computing a weighted sum of values based on query-key similarity rather than treating all inputs equally.

Write the scaled dot-product attention formula used in Transformers.

$$\text{Attention}(Q,K,V) = \text{softmax}\!\left(\frac{QK^{\top}}{\sqrt{d_{k}}}\right)V$$ where $Q$, $K$, $V$ are query, key, value matrices and $d_{k}$ the key dimension.

What architectural innovation defines the Transformer, and why is it advantageous over RNNs?

The Transformer relies entirely on self-attention (no recurrence), enabling full parallelization across sequence positions and better capture of long-range dependencies than sequential RNNs.

Compare TensorFlow and PyTorch in terms of computation graphs.

PyTorch uses dynamic (define-by-run) computation graphs built on the fly, favored for research flexibility; TensorFlow historically used static graphs (define-then-run) for deployment efficiency, though TF 2.x added eager execution.

What are common challenges and techniques in deploying deep learning models to production?

Challenges include latency, model size, and resource limits. Techniques: quantization (reduce numeric precision), pruning (remove redundant weights), knowledge distillation (train a smaller student model), and serving via formats like ONNX or TensorFlow Serving.

In an image classification case study, why is data augmentation used and give examples?

Data augmentation artificially expands the training set and improves generalization/robustness by applying label-preserving transformations such as rotation, flipping, cropping, scaling, and color jittering to input images.