Machine Learning Project: Image Classification Syllabus

What is a Convolutional Neural Network (CNN) and why is it suited to images?

A network using convolutional layers that apply learnable filters across spatial regions; it exploits local connectivity and weight sharing to capture spatial patterns efficiently.

For an input of size $W$, filter size $F$, padding $P$, and stride $S$, give the output spatial dimension of a convolution layer.

$$O = \frac{W - F + 2P}{S} + 1$$

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

It downsamples feature maps to reduce spatial dimensions and computation while providing translation invariance; max pooling keeps the maximum value in each window.

What is transfer learning and why is it standard for image classification with limited data?

Reusing a model pretrained on a large dataset (e.g., ImageNet) and adapting it to a new task; it leverages learned general features, enabling strong accuracy with far less data and compute.

Distinguish feature extraction from fine-tuning in transfer learning.

Feature extraction freezes the pretrained backbone and trains only a new classifier head; fine-tuning additionally unfreezes some backbone layers to update them at a low learning rate.

Compare ResNet and MobileNet at a high level.

ResNet uses residual (skip) connections enabling very deep, high-accuracy networks but is heavier; MobileNet uses depthwise-separable convolutions for a lightweight, fast model suited to mobile/edge with lower accuracy.

What problem do residual (skip) connections in ResNet solve?

They mitigate the vanishing-gradient/degradation problem in very deep networks by letting gradients flow through identity shortcuts, learning $F(x)+x$ instead of $F(x)$ alone.

When compiling the model, what three things must you specify?

The loss function, the optimizer, and the evaluation metric(s).

Which loss is used for multi-class single-label classification, and write its formula for one example.

Categorical cross-entropy: $$L = -\sum_{i=1}^{C} y_i \log(\hat{y}_i)$$ where $y_i$ is the one-hot target and $\hat{y}_i$ the predicted probability for class $i$.

Which activation is used on the output layer for multi-class classification, and write its formula.

Softmax: $$\hat{y}_i = \frac{e^{z_i}}{\sum_{j=1}^{C} e^{z_j}}$$ which turns logits into a probability distribution summing to 1.

Which output activation and loss are used for binary or multi-label classification?

Sigmoid activation per output unit with binary cross-entropy loss: $$L = -[y\log\hat{y} + (1-y)\log(1-\hat{y})]$$

What is the Adam optimizer and why is it popular?

An adaptive optimizer combining momentum and per-parameter adaptive learning rates (RMSProp-style); it converges fast and works well with default settings across many tasks.

Write the basic gradient descent parameter update rule with learning rate $\eta$.

$$\theta \leftarrow \theta - \eta \nabla_{\theta} L$$

What is the learning rate and what happens if it is too high or too low?

The step size for parameter updates. Too high causes divergence or oscillation; too low causes very slow convergence and risk of getting stuck.

Define batch size, epoch, and iteration in training.

Batch size = number of samples per gradient update; epoch = one full pass over the training set; iteration = one gradient update (one batch). Iterations per epoch $= \lceil N / \text{batch size} \rceil$.

What is the ReLU activation and why is it commonly used in hidden layers?

$$\text{ReLU}(x) = \max(0, x)$$ It is cheap, non-saturating for positive inputs, and reduces vanishing gradients compared to sigmoid/tanh.

While monitoring training, how do you diagnose overfitting from the loss curves?

Training loss keeps decreasing while validation loss stops decreasing and starts rising; the train/validation gap widens.

What is early stopping?

A regularization technique that halts training when the validation metric stops improving for a set number of epochs (patience), restoring the best weights to avoid overfitting.

What is dropout and how does it regularize a network?

During training it randomly zeroes a fraction $p$ of unit activations each step, preventing co-adaptation of neurons and acting like training an ensemble, which improves generalization.

What is batch normalization and what benefit does it give during training?

It normalizes layer inputs per mini-batch to zero mean and unit variance (then scales/shifts), stabilizing and speeding training and allowing higher learning rates.

What is a learning-rate schedule, and give one common example.

A rule that changes the learning rate during training; examples include step decay, exponential decay, cosine annealing, and reduce-on-plateau (lower LR when validation stalls).

What distinguishes underfitting from overfitting?

Underfitting = high error on both train and validation (model too simple/undertrained); overfitting = low train error but high validation error (model memorizes noise).

Why do you evaluate on the validation set during training?

To estimate generalization on unseen data, guide hyperparameter tuning and model selection, and detect overfitting without touching the test set.

Define accuracy, precision, and recall using TP, TN, FP, FN.

$\text{Accuracy}=\frac{TP+TN}{TP+TN+FP+FN}$, $\text{Precision}=\frac{TP}{TP+FP}$, $\text{Recall}=\frac{TP}{TP+FN}$.

Write the F1-score formula and state what it balances.

$$F_1 = 2\cdot\frac{\text{Precision}\cdot\text{Recall}}{\text{Precision}+\text{Recall}}$$ It is the harmonic mean of precision and recall, useful for imbalanced classes.

What is a confusion matrix and what does it reveal in multi-class classification?

A table of predicted vs. actual classes; its diagonal shows correct predictions and off-diagonal cells reveal which classes are confused with which.

What does the test set evaluation tell you that the validation set cannot?

An unbiased estimate of real-world generalization, because the test set was never used for fitting or hyperparameter tuning, unlike the validation set.

List three common deployment methods for an image classification model.

Cloud/REST API server (e.g., TensorFlow Serving), on-device/edge deployment (e.g., TensorFlow Lite, Core ML), and batch/offline inference. Browser deployment via TensorFlow.js is another option.

What is model quantization and why is it used for edge deployment?

Reducing weight/activation precision (e.g., float32 to int8) to shrink model size and speed up inference with minimal accuracy loss, enabling deployment on resource-limited devices.

In 'Monitor and Maintain', what is model/data drift and how do you respond?

Drift is when input data distribution or label relationships change over time, degrading accuracy; respond by monitoring live metrics, collecting new labeled data, and retraining/redeploying the model.