Week 1 · Assessment Quiz

Deep Learning Intro & FastAI

25 multiple-choice questions on deep learning fundamentals and FastAI basics, plus 5 short-answer questions.

📋 30 questions total ⭐ 30 marks 🕐 No time limit 🔒 Answers not revealed

Deep learning is best described as:

AA set of hand-crafted feature extraction rules BA branch of machine learning using multi-layered neural networks that learn representations automatically CA statistical technique limited to linear models DA rule-based expert system that encodes domain knowledge

In FastAI, what does learn.fine_tune(3) do?

ATrains only the final classification layer for 3 epochs BFreezes the base model, trains the head for 1 epoch, then unfreezes and trains all layers for 3 more epochs CTrains the entire network from random initialisation for 3 epochs DEvaluates the model on the validation set 3 times

Transfer learning in image classification involves:

ATraining a model from random weights on your dataset BUsing weights pre-trained on a large dataset as a starting point for a new task CCopying the architecture of one model to another without any weights DTraining multiple models and averaging their predictions

The purpose of a validation set during training is to:

AProvide more data for gradient updates BSpeed up training by using smaller batches CEvaluate model performance on unseen data to detect overfitting DReplace the test set at deployment time

The Oxford-IIIT Pet dataset is used in Week 1 to demonstrate:

AText sentiment classification BMulti-class image classification of cat and dog breeds CObject detection with bounding boxes DImage segmentation

In FastAI, error_rate is defined as:

A1 minus accuracy (fraction of incorrectly classified samples) BThe total loss divided by the number of samples CThe standard deviation of the loss over the validation set DThe number of misclassified images in the training set

Which of the following is NOT a type of layer found in a standard convolutional neural network?

AConvolutional layer BPooling layer CFully connected (dense) layer DLogistic regression layer

A DataLoader in FastAI is responsible for:

ADownloading datasets from the internet BBatching, shuffling, and feeding data to the model during training CSaving model weights to disk DPlotting training metrics

vision_learner(dls, resnet34, metrics=error_rate) creates a learner that:

ATrains a ResNet-34 from random weights on the given data BUses a pre-trained ResNet-34 backbone with a new head suited for the number of classes in dls CLoads a ResNet-34 and freezes all layers permanently DDownloads ResNet-34 weights and evaluates on the validation set only

Q10

When we say a model is "pre-trained", we mean:

AThe model has been reviewed by an expert before use BThe model was trained on a very small dataset as a warm-up step CIts weights were learned on a large dataset (e.g. ImageNet) before being adapted to our task DThe model architecture was designed specifically for our dataset

Q11

The loss function during training measures:

AHow long it takes to complete one epoch BHow far the model's predictions are from the true labels CThe number of parameters in the model DThe learning rate schedule used by the optimiser

Q12

A mini-batch in deep learning refers to:

AA lightweight version of a neural network BA single training example passed through the network CA small subset of the training data used for one gradient update step DThe smallest possible learning rate

Q13

The sigmoid activation function is most appropriate for:

AThe output layer of a binary classification model BHidden layers in deep convolutional networks CThe input normalisation step DComputing attention weights in Transformers

Q14

If the learning rate is set too high, training will likely:

AConverge extremely slowly but reach a better minimum BDiverge — the loss oscillates or explodes rather than decreasing steadily COverfit the training data in exactly one epoch DHave no effect because the optimiser self-corrects

Q15

learn.predict(img) in FastAI returns:

AThe raw logits before any activation function BThe training loss for that image CThe predicted class label, its index, and the probability tensor DOnly the index of the predicted class

Q16

A confusion matrix for a 3-class classifier has:

A3 cells total (one per class) B6 cells total C9 cells (3 × 3) DThe number of cells depends on the dataset size

Q17

Data augmentation (e.g. random flipping, cropping) helps because:

AIt increases GPU memory efficiency BIt artificially expands the effective size of the training set and improves generalisation CIt speeds up model inference at test time DIt removes noisy labels from the training data

Q18

In FastAI, ImageDataLoaders contains:

AOnly the training dataset BOnly the validation dataset CBoth a training and a validation DataLoader, along with vocabulary and transformations DThe test dataset used for final evaluation

Q19

Fine-tuning a pre-trained model means:

AStarting training from completely random weights BContinuing training on a new task, adapting the pre-trained weights rather than discarding them CFreezing all layers and only predicting with the model DReducing the model size by removing layers

Q20

ResNet (Residual Network) is notable for introducing:

ASkip connections that allow gradients to flow past layers, enabling very deep networks BAttention mechanisms between encoder and decoder CDepthwise separable convolutions for efficiency DBatch normalisation as the primary training technique

Q21

GPUs are used for deep learning training primarily because:

AThey have more RAM than CPUs BThey are massively parallel processors suited to the matrix operations that dominate neural network computation CThey run at a higher clock speed than CPUs DThey are cheaper than CPUs for cloud computing

Q22

learn.lr_find() in FastAI helps you:

AAutomatically set the learning rate to an optimal value without any user input BFind a good range of learning rates by training briefly and plotting loss versus learning rate CFind the minimum loss achievable on the validation set DSelect the best model architecture for your dataset

Q23

Which metric is most appropriate for an imbalanced binary classification problem?

AAccuracy alone BError rate alone CPrecision, Recall, or F1-score, which account for class imbalance DMean squared error

Q24

One complete pass through the entire training dataset is called:

AA batch BAn epoch CA learning rate schedule DA gradient step

Q25

The Oxford-IIIT Pet dataset challenge is that:

AThe images have very low resolution making features hard to detect BThere are 37 categories with similar-looking breeds, requiring fine-grained visual discrimination CThe dataset has severe class imbalance between cats and dogs DLabels are provided as bounding boxes rather than class names

Answer each question in 2–4 sentences. Precise technical language is expected. Code snippets are welcome where relevant.

Q26

What is the difference between learn.fit() and learn.fine_tune() in FastAI?written

Your answer

Think about which layers are frozen and for how long.

0 / 700

Q27

Explain what transfer learning is and why it is useful when working with small datasets.written

Your answer

0 / 700

Q28

What is a confusion matrix? What can it tell you about model errors that overall accuracy cannot?written

Your answer

0 / 700

Q29

Describe the role of the validation set during training. Why is it important to keep it completely separate from the training set?written

Your answer

0 / 700

Q30

List and briefly explain the main steps to build an image classifier in FastAI, from loading data to evaluating performance.written

Your answer

0 / 700

Your full name

Complete all 30 questions then click Submit. Your MCQ score (25/25) will be shown. Short answers are marked separately.

Deep Learning Intro & FastAI

Multiple Choice (25 marks)

Short Answer (5 marks — marked by lecturer)